The AI Hype Audit

The claim15-second reel, ~187 reactions. On-screen cards headed '10K/MONTH' walk through four assets: a POSTING SCHEDULE (Reels 1x/day, photo posts 2-4/week, carousels 3/week, Stories 1-3x/day), a DM SCRIPT to 'nurture followers into those calls', a PRE-CALL SCRIPT (send a confirmation thank-you page, send a video sales letter, send a pre-call video) with the script shown on screen, and a SALES SCRIPT for the call itself. Closes: 'if you want all this broken down into a cheat sheet so you can follow along and use it, [comment] and I'll send it over.'

What is real: every mechanic named is standard, unremarkable direct-response practice that legitimate agencies do use. A confirmation page after booking, a video sales letter, and a pre-call video are the ordinary 'show-up rate' stack, and they genuinely do lift attendance on booked calls. A daily-reels cadence with carousels and Stories is a normal organic content posture. Nothing shown is fabricated software or a fake feature. What is not real is the number on the front of it. '10K/MONTH' is presented as the output of the posting schedule, and nothing in the reel connects the two: no offer price, no close rate, no call volume, no time period, no accounting for ad spend or the coach's own audience size, and no evidence that anyone following this schedule earned anything. An income figure with no denominator is not a result, it is a headline. Note the contrast with entry 275 in this same batch, where a company making a similar pitch put 'Information not typical. These are examples not guarantees. Results may vary.' on screen. This reel carries no such disclosure. The ask is also the tell: the entire piece is engineered to make you comment a keyword, which is a lead-capture mechanic, not an act of generosity. The cheat sheet is the top of a funnel that ends somewhere paid.

The four assets are ordinary, legitimate coaching-funnel mechanics worth building. The $10,000 a month attached to them is an unsubstantiated headline with no disclaimer, and the reel's real purpose is to collect your comment.

Confidence

High

Posted by

A fitness-styled business coach (reel sent to the self-thread Mon Sep 7, 2026, 7:53 AM)

See the original claim →

The claim11-second screen-recording reel, ~83.4K reactions, working in a PSD titled 'adobe photoshop-shortcut-KEYS'. Three shortcuts are demonstrated and labeled: ALT+CTRL+G = 'CLIPPING MASK' (a portrait clipped into a cluster of circles), ALT+SHIFT+S = 'BLEND FOR BLACK' (a flame on a black background dropped onto a fire-pit photo so the black disappears), and ALT+SHIFT+M = 'BLEND FOR WHITE' (birds on a white background dropped onto a sunset so the white disappears).

All three are correct, and the two blend-mode ones are the kind of shortcut most people who have used Photoshop for years still do not know. Alt+Ctrl+G is genuinely Layer > Create Clipping Mask. Shift+Alt+S sets the active layer's blend mode to Screen, and Screen does exactly what the reel shows: it drops black to transparent, which is why it works on flames, fireworks, smoke and light leaks shot against black. Shift+Alt+M sets Multiply, which drops white, which is why it works on line art, birds, signatures and anything scanned on paper. The letter mnemonics hold up (S for Screen, M for Multiply), and the rest of the family works the same way (Shift+Alt+N for Normal, Shift+Alt+O for Overlay). The one thing the reel leaves out matters in practice: these blend-mode shortcuts only change the LAYER's blend mode when the Move tool, or another non-painting tool, is active. If you have the Brush, Clone Stamp or Eraser selected, the same keystroke silently changes that TOOL's blend mode instead and your layer does not change, which is the exact moment a beginner concludes the shortcut is broken. Press V first and it works every time. Also worth saying plainly: Screen and Multiply are not a substitute for a real cutout. They work because the background is pure black or pure white, and they will wash out or darken anything that is not.

All three shortcuts are correct and genuinely useful. Press V first, and remember Screen and Multiply only work on pure black or pure white backgrounds.

Confidence

High

Posted by

@kkpictures88 (Photoshop tips account; reel sent to the self-thread Mon Sep 7, 2026)

See the original claim →

The claim2 minute 33 second reel, ~81 reactions, sent to the self-thread Sun Sep 7, 2026. Billed as the '80/20 LANDING PAGE METHOD'. Walks through a live teardown of the company's own scaling-workshop page: lead with the headline as the question the visitor already has ('Are you the biggest risk to your own business?'), a one-line description under it, a form that asks for as little as possible ('Get your free personalized $100M scaling roadmap... in under 30 seconds'), an objections block answering the questions that would otherwise stop the signup, and a fast mobile experience, all aimed at 'qualified leads.' Ends on a promotion for a new marketing workshop.

There is no false factual claim here to knock down. Everything taught is standard, decades-old direct-response landing-page practice: one page, one job; the headline carries the offer; strip the form to the minimum fields; answer objections on the page rather than hoping the prospect emails you; make it fast, because mobile abandonment is real and measurable. This is the same advice in every serious conversion-rate book, and the person teaching it is demonstrating it on a page his own company actually runs, which is the honest version of a teardown. Two things earn it the 'real' verdict rather than 'partly.' First, the advice is not gated behind a fabricated mechanism or a tool that does not exist. Second, and this is rare enough in this corpus to be worth naming, the closing frames carry an on-screen disclosure: 'Information not typical. These are examples not guarantees. Results may vary.' Almost nothing else audited this month does that. Be clear-eyed about what it is, though: it is content marketing. The reel exists to sell a marketing workshop, the free '$100M scaling roadmap' is a lead magnet, and the page being torn down is the same page collecting your details. That is a funnel. It is just an honestly labeled one, teaching real technique.

Genuinely good landing-page advice, demonstrated on the company's own live page, and one of very few pieces in this corpus to display an earnings disclaimer. It is still an ad for a workshop.

Confidence

High

Posted by

Acquisition.com (the on-screen pages, forms and workshop branding are all Acquisition.com's own)

See the original claim →

The claim13-second screen-recording reel working in a file named 'adobe illustrator Batman.ai'. Three shortcuts demonstrated: Ctrl+8 labeled 'COMPOUND PATH' (a bat silhouette knocked out of a red sphere to make the logo), Ctrl+7 labeled 'CLIPPING MASK' (a Batman image masked into a set of vertical bars), and Ctrl+Alt+C, shown but never labeled, warping the word BATMAN to fit inside the bat-wing shape.

Correct on all three. Ctrl+8 is Object > Compound Path > Make, and using it to knock one shape out of another is exactly what it is for. Ctrl+7 is Object > Clipping Mask > Make, and the topmost object becoming the window is the correct behaviour shown. The third one is the interesting one and the reel never names it: Alt+Ctrl+C is Object > Envelope Distort > Make with Top Object, which takes the shape on top and uses it as a mould for everything beneath. That is why the BATMAN lettering ends up stretched into the wing silhouette. It is a real command, it is the right shortcut, and it is the most useful of the three for anyone laying out a logo, which makes it odd that it is the only one left unlabeled. Two caveats the reel does not give you. A compound path takes the appearance of the bottom-most object, so if the knockout comes out the wrong colour, that is why, not a bug. And envelope distort leaves the text live and editable until you go Object > Expand, which you will want to do before sending artwork to a printer, because envelopes are a common source of RIP failures. Worth noting for anyone tempted to follow along with these exact files: the Batman logo is DC Comics intellectual property. Fine as a practice exercise on your own machine, not fine on anything you sell.

All three shortcuts are correct. Alt+Ctrl+C, the one left unlabeled, is the one worth learning. Expand envelopes before print.

Confidence

High

Posted by

@kkpictures (Adobe tips account, same family as entry 276; reel sent to the self-thread Sun Sep 7, 2026)

See the original claim →

The claimImage post sent to the self-thread Sun Sep 6, 2026, 10:26 PM. The graphic reads 'GPT-6 ASTRA - What it is. What it isn't. Why it changes everything' over 'IS THIS THE END OF CLAUDE? ARE WE IN THE AGI ERA?' The body opens: 'Is this the end of Claude? Absolutely not.' The author states he is a heavy Claude user who probably uses it more than ChatGPT, praises Anthropic, and says this is 'not a Claude is dead post.' His actual argument: OpenAI's GPT-6 Astra raises the question of how fast the whole category is changing, as frontier systems move from 'ask me a question' to 'give me the objective, I'll help execute the work.'

The factual core checks out, and that is worth stating up front because the instinct is to assume a post like this is invented. It is not. OpenAI announced GPT-6 Astra on September 3, 2026, four days before this post was sent, initially as a limited preview for trusted partners, rolling out to ChatGPT Plus, Pro, Business and Enterprise plans as well as the API and AWS. It was reported as scoring at or near the top of key reasoning benchmarks, ahead of both OpenAI's previous GPT-5.6 Sol and Anthropic's Claude Fable 5, and it is OpenAI's first model to reach the Critical cybersecurity tier under their Preparedness Framework, with companies in their application-based cybersecurity programme getting first access. So the release is real, the name is real, the date is real, and the benchmark framing is roughly as reported. The author's actual thesis, that the competition is shifting from answer quality to executing whole objectives, is a reasonable read and he argues it honestly. What drops this to 'partly' is the packaging. The graphic asks 'Is this the end of Claude?' and 'Are we in the AGI era?', and the very first line of his own body copy answers the first question with 'Absolutely not.' That is a headline written to be clicked, walked back one sentence later, which is precisely the pattern this audit exists to name, even when a thoughtful post sits behind it. 'Are we in the AGI era?' goes unaddressed entirely; a benchmark score is not a definition of AGI, and no benchmark result settles that question. The graphic also carries aimasterclub.com branding, so the post is doing promotional work alongside the commentary.

The underlying news is real and verified: GPT-6 Astra released September 3, 2026, benchmarking ahead of Claude Fable 5. The clickbait headline is contradicted by the author's own opening line, and the AGI question is never answered.

Confidence

High

Posted by

Zeev Wexler (verified page; the attached graphic is branded AI Master Club, aimasterclub.com)

See the original claim →

The claim24-second reel, ~62K views, sent Sun Sep 6, 2026: 'This 1 Claude skill fully replaces your Higgsfield subscription.' Kie.ai's pay-as-you-go API plus a Claude skill makes 'the same viral AI videos' for about $0.75 each with no monthly bill, 'sooo much more affordable.' On screen: a table reading '$1 gets you 1 finished 10-second clip at 1080p' vs 'Nothing. There is no $1 plan,' '5 clips: $3.75' vs 'Higgsfield $15 a month at 720p and 8 seconds,' credits 'carry over forever' vs 'reset to zero every month,' and a faceless TikTok channel shown as proof (260.5K followers). Comment 'LAVENDER' for the tutorial.

What is real: Claude Skills exist (Anthropic, Oct 2025). The skill is real and free; it is the 'generate' skill in the Blotato-Inc/blotato-skills GitHub repo (154 stars), which drafts a still, animates it through Kie.ai, then 'hands off to Blotato.' Kie.ai is real: NexusAI Services LLC, a Colorado LLC formed May 2024, reselling third-party models (Veo, Kling, Seedance, Sora, Runway) at $0.005 a credit with a $5 minimum top-up and credits that do not expire. Her Aug 19, 2026 Substack post lays out the five steps and the roughly $0.75 per 10-second 1080p figure, which is in range for Kie's Seedance and Kling-standard pricing ($0.057 to $0.07 a second). What is not: 'Same videos' is a stretch. Higgsfield's Plus and Ultra tiers bundle 'unlimited' Seedance 2.0 and Kling 3.0 in the web app plus studio exclusives Kie does not carry. 'Sooo much more affordable' only holds at low or irregular volume: her own company's blog admits that at matched 1080p Seedance 2.0, Kie runs 'roughly two to four times the per-clip cost' of a fully used Higgsfield Plus plan. The $1-vs-$15 table is rhetoric: there is no $1 Kie top-up (minimum $5), and Higgsfield's entry plan has flipped between $9, $15 and $19 at least three times in 2026, so '720p, 8 seconds' is one snapshot. The undisclosed part: Blotato is her company. Every link in the tutorial carries a Blotato ref tag, the skill's last step is Blotato ($29 to $499 a month), and Blotato pays affiliates 50 percent of a first payment. No Kie.ai affiliate program was found, and no Higgsfield affiliate link is in play, so the steer is toward her product, not a commission. Kie's Trustpilot sits at 2.2 of 5 with complaints of credits eaten by failed generations. Honest route: pay Kie directly if you post a few clips a month; go Higgsfield Plus if you post daily; skip the Blotato step unless you already need a scheduler.

You can produce a 10-second 1080p AI clip for about $0.60 to $0.75 through Kie with a free Claude skill, but 'replacing Higgsfield' only saves money at low volume, and the funnel ends at her own $29-plus a month product.

Confidence

High

Posted by

Sabrina Ramonov (AI creator with a large following; founder of Blotato)

See the original claim →

The claim30-second reel, ~34K views, sent Sun Sep 6, 2026, titled 'How to remove Claude watermarks in 30 seconds.' Caption: 'Pass all AI detectors with this. Comment "water" for the full guide.' The screen recording shows the GitHub repo guillaumemeyer/watermarks-remover and its remove-ai-marks skill running inside Claude: Layer A strips hidden Unicode and EXIF/C2PA metadata, Layer B does a 'humanize pass' that rewrites the wording. Claude's own on-screen output reads: 'Nobody can prove "undetectable," and I won't claim it.'

The repo is real and, ironically, honest. guillaumemeyer/watermarks-remover (MIT, created Aug 11, 2026, ~20.9K stars, 2.4K forks, v0.7.0, pushed Sep 5) does two things: Layer A deterministically strips invisible Unicode and C2PA/EXIF/XMP metadata from 20-plus file types, and Layer B rewrites prose to try to break statistical watermarks. Its own docs say Layer B 'cannot claim official undetectable,' that 'no tool can honestly certify this fails the official check,' and that 'rewording degrades the copy.' Its ethics file lists academic fraud and circumventing disclosure rules as misuse. The caption promise is the BS. Anthropic's text watermark (default-on, worldwide, every model launched after Aug 2, 2026; see #181) is a SynthID-style pattern in word choices, not a hidden character, so Layer A does nothing to it. Layer B is the only attempt, and the skill itself says to run the rewrite with a model that is NOT the suspected origin, because output from Claude gets watermarked again. The reel shows Claude doing the humanize pass. That is the self-defeating loop from #181: you launder the mark by asking the marker to re-mark it. Nobody can prove 'passes all detectors' anyway: Anthropic's detection API is still unreleased, so there is nothing official to test against, and the commercial detectors disagree with each other. GPTZero claims 93.5 percent recall on paraphrased text, Turnitin ships a dedicated 'bypasser' flag, Originality claims 96.7 percent on paraphrased text, and independent tests show some humanizers slip through while some human writing gets flagged. Unreliable in both directions means 'pass all' is not a product anyone can sell. The reel is a lead magnet: 'water' feeds a DM into a guides site and a $59 a month Skool community. The honest route: strip your own EXIF if you want, use AI as a drafter, edit it for real, and disclose where disclosure is required.

You can clean metadata in 30 seconds; you cannot buy 'undetectable,' and the one detector that matters for Claude text isn't even public yet.

Confidence

High

Posted by

Tenfoldmarc (AI-tips Facebook page)

See the original claim →

The claim10-second reel, ~9.6K views, sent Sun Sep 6, 2026: 'Spend 30 minutes on this Claude setup and every answer stops sounding generic.' Recipe in the caption: open Cowork instead of chat, turn Extended Thinking on, select Opus 4.8, point Claude at a 'Claude Work' folder with About Me / Projects / Templates / Outputs, add about-me.md, my-voice.md and my-rules.md, then paste a five-line Global Instruction ('Read my files before every task. Ask questions before executing. Show a plan first. Never delete without approval.'). Comment 'CLAUDE' for the poster's 'Ultimate Claude Guide.'

The mechanics check out. Claude Cowork is a real mode in the Claude desktop app (Mac, Windows, and since July 7, 2026 web and mobile) on Pro, Max, Team and Enterprise. Settings > Cowork > Edit Global instructions exists, and Anthropic's own Cowork course tells you to put exactly this kind of thing there: who you are, your shorthand, how you want output delivered. Cowork can be pointed at a local folder, and Cowork projects hold instructions, context folders and memory. Keeping plain markdown files about yourself and telling Claude to read them is a legitimate, documented pattern; Claude Code's CLAUDE.md is the formal version of it. What is off: the script is a near word-for-word lift of Ruben Hassid's free March 6, 2026 post (same folder names, same three files, same five-line instruction) with 'Opus 4.6' swapped for 'Opus 4.8.' Opus 4.8 exists (May 28, 2026) but Anthropic lists it as legacy; Opus 5 replaced it July 24 and Fable 5.1 sits above that since September 1. On the current models the switch is labelled 'Thinking' under Effort, and on Opus 5 and Fable 5.1 it cannot be turned off, so 'turn Extended Thinking on' is a step from spring. Nothing in Anthropic's docs says Cowork auto-reads about-me.md; the 'Read my files before every task' line is what does that work, and it is context, not enforcement. 'Sounds like you every time' is oversold; Anthropic's own docs say instruction files are followed, not guaranteed. The funnel: no Menler website, store or guide could be found anywhere public beyond the menler.in card shown in the reel. 'Comment CLAUDE' is a DM-capture mechanic; the likeliest payload is the free Hassid material, which itself gates downloads behind a Substack signup. Honest route: do the setup from Anthropic's free course, skip the comment.

Expect noticeably less generic drafts within a week if my-voice.md contains real samples, with occasional drift you correct by adding a line to Global instructions; nobody gets 'sounds like you every time.'

Confidence

High

Posted by

Menler (AI-tips Facebook page, 'India's only Claude-first learning brand')

See the original claim →

The claimExplee AI's Facebook page posted Aug 27, 2026: 'Drop your website link. Our AI finds the people who need your product, writes the emails and books the calls. You get $30 in credits to try it. About 10 first leads, no card games.' The creative is a screenshot styled as an X post from 'Sam' @samships (blue check, 402K views, 4.2K likes, timestamped Aug 3, 2026) reading 'i'm sorry, Claude can do WHAT?! it finds your buyers, writes the emails and books the calls. while you build. $30 free to try,' with a reply from 'Maya' @mayabuilds. Link card: 'EXPLEE.COM — Get your first leads free.' Sent Sat Sep 5, 2026.

The product under the ad is real and we have graded it three times before (#036, #054, #196): Explee LTD (UK company 15759064, incorporated June 2024, London) sells AutoGTM, a pay-as-you-go cold-email agent at about $0.03 an email with $30 in starter credits. Its own pricing page says $30 buys roughly 1,000 emails, '~2 to 8 warm leads' and '~1 to 2 meetings.' That is where the honest part ends. The social proof is fabricated. A fresh fetch on Sep 6 shows @samships is a protected fan account from 2017 with 6 followers, 7 tweets and no verification, and @mayabuilds is a Roblox clothing designer with 0 followers. Neither could have posted a 402K-view, 4.2K-like tweet on Aug 3, 2026, and the text exists nowhere on X or the web. The same '402K views' prop appeared in the tweet screenshot we graded BS at #054 on Aug 9; this is the fourth Explee ad in the archive and the second with an invented tweet. 'Claude' is borrowed credibility: a co-founder's June 12 launch post says AutoGTM is 'powered by Fable 5' (an Anthropic model bought through the API), but explee.com never mentions Claude and there is no partnership. 'No card games' is false for the agent: the AutoGTM terms (updated July 12, 2026) require a saved payment method before the agent starts and authorize off-session charges with retries; a Trustpilot reviewer this week says there is no way to cancel. 'About 10 first leads' exceeds Explee's own 2 to 8 range. 'Books the calls' means the bot replies with an invite when a prospect says yes; you still hold the call, handle objections and close. Honest route for a small business: warm intros, a complete Google Business Profile, referral asks to existing customers, and if you test cold email, do it under your own domain with CAN-SPAM basics at 30 emails a day, not 1,000 for $30.

A small business that spends the $30 will most likely see a few replies, zero to two calls, and a saved card that keeps charging unless it is removed; the 402K-view 'Sam' will never exist.

Confidence

High

Posted by

Explee AI (company Facebook page, paid-ad style post)

See the original claim →

The claim23-second reel, caption '$40k in free traffic from Google (every month)' over a Semrush-style organic chart reading '14.9K, +1,388%' and a note: "Find a question (FindQuestions.com). Answer it using ChatGPT, prompt in the comments. Edit + add images. 3 internal links. Post it." Spoken line: "It's crazy how easy this is." Sep 5, 2026.

The recipe is ordinary, legitimate SEO: find a real question your customers ask, answer it properly on your own site, add images, link it internally. What the reel sells on top of that is a number that is not money. '$40k' matches the 'traffic cost' style metric SEO tools print next to an organic-traffic estimate: what that traffic would cost to buy as Google Ads. Semrush states its traffic figures are accumulated and approximated from clickstream panels, not measured from the site. So the screenshot shows an estimate of an estimate; nobody's bank account moved $40,000. The funnel is the second half: FindQuestions.com is HeyTony's own tool. It surfaces Reddit questions by business type, gives a free preview, and upsells 'HeyTony Insiders' at $97 a month, and its own disclaimer says results are not guaranteed to drive traffic or rankings. The risk the reel never mentions: Google's spam policy bans 'scaled content abuse', which it defines as using generative AI to produce many pages without adding value, and its March 5, 2024 core update specifically went after pages that claim to answer popular queries but deliver minimal value, reporting 45 percent less low-quality content in results by that April. AI-assisted pages are allowed when they are people-first and original. Done carefully, two to four real question pages a month is good practice; done as a volume play, it is the exact thing Google is filtering out.

Two to four honest question pages a month is the version of this that works for a local business; the $40k belongs to the tool that printed it.

Confidence

High

Posted by

HeyTony (Matt Diamante, Hamilton ON SEO agency)

See the original claim →

The claim20-second screen-recorded reel captioned 'Now you can use AI chatbot for free' showing uncensored.ai: a '3/5' free-plan counter, a 'You've reached your daily limit of 5 responses, Upgrade to Pro' wall, a settings page with 'Model Tone: Analytical / Direct / Crude' and a 'Clean Language' toggle, and the bot refusing 'how to hack the phone' and 'how to make a bomb in 50 USD' while writing a spy story on request. Sep 5, 2026; 6.3K reactions.

The product exists and the free tier is real: Uncensored.AI's pricing page offers up to five responses per day at no cost, then Pro at $9.99 a month or $99.99 a year. The two words in the pitch that matter, 'free' and 'uncensored', are both weaker than they sound. Free is table stakes in 2026: ChatGPT, Claude and Gemini all run free tiers with browsing, images, files and far more than five messages a day. And 'uncensored' is marketing: its own terms of service prohibit illegal or harmful use, which is exactly why the reel shows it declining the hacking and bomb prompts like any mainstream assistant would. Under the hood it does not run its own model; its privacy policy says prompts are sent to Parasail, an inference host for open-weight models, and it does not disclose which model answers you. The registrant is a small Nebraska LLC; ScamAdviser rates the domain 'very likely safe' but flags a Gmail contact address and possible adult content. If a business owner wants a free assistant, the real question is not price but data: Gemini's default activity setting can use chats for training and human review unless turned off, and Claude's training is opt-in; that is the setting to check before pasting anything about customers.

A real product with a real free tier, oversold on both of its adjectives.

Confidence

High

Posted by

Piyushikaamriti

See the original claim →

The claimText post: '🚨 BREAKING: CHATGPT can now create a full YouTube channel for you with AI, like an agency that generates $10,000 a month. Free. Here are 7 prompts.' Sep 5, 2026.

Three claims stacked, none of them true as stated. First, ChatGPT does not create a YouTube channel: it can write scripts, titles, descriptions and outlines, and the free tier does that with usage caps, but it does not generate the videos, voice them, make thumbnails at scale, or upload and publish anything. OpenAI's own video model, Sora, was discontinued in spring 2026 according to Wikipedia's Sora and OpenAI articles as fetched Sep 5 2026 (announcement March 24, app shutdown April 26, API to follow Sep 24), and even while it existed it was for paying Plus and Pro users only. Video today means paid tools: Google's Veo, for example, sits behind the $19.99 and $99.99 Gemini plans. Second, 'free' ignores that YouTube pays nothing until a channel clears the Partner Program bar of 1,000 subscribers plus 4,000 public watch hours in a year, or 10 million Shorts views in 90 days. Third, '$10,000 a month' at typical RPMs of $1 to $30 per thousand views, with creators keeping 55 percent of long-form ad revenue and 45 percent on Shorts, needs on the order of one to two million monthly views; that is an outlier outcome, not what a prompt pack produces. And YouTube's July 15, 2025 inauthentic-content rule demonetizes AI-generated content made with generic or unoriginal templates that gives the impression of mass production, with undisclosed realistic AI content risking removal or Partner Program suspension. The '7 prompts' are the hook for the page's own funnel; nothing in them changes any of the above.

The prompts are free; the channel, the video, the audience and the $10,000 are not, and ChatGPT does not produce any of them.

Confidence

High

Posted by

Daniel Design (a page)

See the original claim →

The claim24-second reel over the OpenMAIC GitHub README (29k stars): "Turn PDFs into AI classrooms… AI will answer anything. It won't teach you… Drop in any topic/pdf… Includes: quizzes, interactive AI teacher + students… Students don't watch. They pull the levers… Teacher can call on you… Self-host the whole thing, bring your own models… Comment 'CLASS' and I'll DM you this repo… Grab it or pay later." Sep 5, 2026.

The tool in the reel exists and does what the captions say. THU-MAIC/OpenMAIC is a public GitHub repo created Mar 11 2026, showing 31,763 stars, 5,220 forks and an MIT license as of Sep 5 2026 (the reel's '29k' is simply a few days old; last push Sep 4 matches the '13 hours ago' commits). v1.0.0 shipped Aug 27 2026, adding a chat-driven 'Pro workbench' on top of the classic one-click generator. THU-MAIC is Tsinghua University's MAIC lab; the underlying research is 'From MOOC to MAIC' (arXiv 2409.03512, Sep 5 2024; JCST vol. 41 no. 1, Jan 2026, 500+ students, 100k+ learning records) and MAIC-UI (arXiv 2604.25806, Apr 28 2026). The README confirms the pipeline (Outline stage then Scenes stage), inputs (topic, PDF, Word, PPTX, spreadsheets, text, images, audio, video), exports (.pptx, interactive HTML, classroom ZIP, MP4 via an optional Chromium/FFmpeg render container), providers (OpenAI, Anthropic, Azure, Gemini, DeepSeek, Ollama and more) and OpenClaw chat integration (Feishu, Slack, Telegram, Discord, WhatsApp). Self-hosting needs Node 22.19+, pnpm 10+, one LLM key; Postgres is optional. A free hosted demo runs at open.maic.chat. 'Comment CLASS and I'll DM you' is a standard ManyChat-style comment-keyword auto-DM (replient guide, updated Apr 23 2026) for list-building; the repo is one search away. 'Grab it or pay later' has no receipt: the project moved from AGPL to MIT on Jun 29 2026 (v0.3.0), the opposite direction of going paid.

A small business can realistically turn an existing PDF manual into a reviewed, narrated training module for a couple of dollars in API spend, but the reviewing step is not optional.

Confidence

High

Posted by

a creator page

See the original claim →

The claimReel: "Amazing Websites for UI/UX Designers" — 1. shaders.com ('the component library for creative WebGPU effects in the browser'), 2. ls.graphics ('high-end mockups for serious design work'), 3. ContentCore.xyz ('3D mockups and 3D motion templates'). Sep 5, 2026.

Frames confirm the reel names exactly three sites, and all three exist and are live as of Sep 5 2026. (1) shaders.com is run by Shader Effects, Inc. (Delaware C-Corp); the npm package 'shaders' (v3.2.466, published Sep 3 2026) renders WebGPU effects for Vue/React/Svelte/Solid. The on-screen line 'component library' implies open source; it is not. The license (effective Aug 3 2026) says free use is evaluation only and 'any commercial or production deployment' needs Core ($59/yr) or Pro ($199/yr); SaaS use needs a separate OEM deal. It is WebGPU-only with 'no WebGL fallback', so it will not render on older browsers. Do not confuse it with paper-design/shaders (Apache-2.0, WebGL, Lost Coast Labs). (2) ls.graphics ('Lovely and Spotless Graphics') is a long-running mockup shop: Start $9/mo or $99/yr (online editor, PNG/JPG), Pro $12/mo or $144/yr (PSD/Sketch/Figma downloads, Figma plugin, 1.2M+ installs), 14-day refund, a small free-mockups section. 'High-end' is fair; nothing on screen says paid. (3) ContentCore.xyz is a browser-based 3D mockup and motion tool (Awwwards Honorable Mention Feb 16 2025; enterprise via genbrand.design; no named founder). Basic is free with unlimited 4K image and 720p video renders but 'non-commercial use only'; Commercial is $9.99/mo or $5.99/mo billed annually ($71.88/yr). The list is accurate, just silent on cost and license; a business owner can use only ContentCore's free tier, and only for non-commercial output.

Expect nicer visuals for $6-$12/mo per tool, not a free upgrade; the only thing free here is looking.

Confidence

High

Posted by

a design-creator page

See the original claim →

The claimReel: "Client: I don't want basic website / Me: How's this?" over a bioluminescent 'Enter the Rabbit Hole' 3D scroll site. Caption: "Build this Custom 3D scroll Website — Follow and Comment 'RABBIT' for the prompt AND workflow — DM me to get a modern website built for your business!" Sep 5, 2026.

The reel is a textbook comment-to-DM funnel. ManyChat's own guide (published Sep 16 2024, updated Jul 6 2026) describes it exactly: a viewer comments a keyword, a bot DMs 'your link, your offer, or your freebie in seconds,' optionally gated behind a follow. The caption's real ask is the last line: 'DM me to get a modern website built for your business.' The 'prompt and workflow' is the lead magnet; the developer's paid build is the product. Meta's Transparency Center classes posts that 'explicitly request engagement (such as votes, shares, comments...)' as engagement bait, and a 1kreach analysis (Apr 26 2026) calls the 'comment X to get the link' DM funnel the most penalised pattern on Reels (though it cites no data). On the tech: scroll-driven 3D sites are real and well documented. Codrops (Nov 19 2025) shows the stack (Three.js/R3F, GSAP ScrollTrigger, GLB models, custom shaders) and rates it intermediate-to-advanced. A 2026 developer guide puts a basic version at 2-4 hours with prerequisites, a commercial build at 16-40 hours, and self-learning to commercial quality at 6-12 months. Claude Code can scaffold Three.js+GSAP scroll pages (MindStudio, Mar 17 2026: 'an afternoon, under $10', but with 3-4 revision rounds and a warning that video+Three.js+scroll animations 'equals slow page on mid-range devices'). No single prompt hands you a finished, hosted, branded 3D site with your own model. Agency pricing for a focused 3D hero site: $20K-$50K, 8-12 weeks (Utsubo cost guide). And Utsubo's May 20 2026 SEO guide notes an empty canvas 'for the first three seconds' makes LCP three seconds; their fix was a static image on mobile.

Expect a bot DM with a link and a pitch; the technique is real, the 'one prompt' is not, and a 3D scroll site is usually the wrong spend for a local business.

Confidence

Medium

Posted by

a web-dev creator page

See the original claim →

The claimStatic infographic reel: "20 Must-Have Claude Connectors" — Gmail, Slack, Google Drive, Notion, Microsoft 365, Google Calendar, Granola, HubSpot, Apollo.io, Canva, Gamma, Figma, MailerLite, Kit, Asana, Linear, GitHub, Jira, Zapier, n8n — each with a one-line job (Canva 'generate and edit designs from a prompt', Figma 'audit design systems, generate FigJam diagrams', n8n 'trigger automations across your stack via chat'). Sep 5, 2026.

Checked against claude.com/connectors on Sep 5, 2026: Gmail, Slack, Google Drive, Notion, Microsoft 365, Google Calendar, Granola, HubSpot, Apollo.io (slug /connectors/apollo), Canva, Gamma, Figma, MailerLite, Asana, Linear, GitHub, Zapier and n8n all have official listing pages; Jira ships as the 'Atlassian' (Jira + Confluence) connector. The one miss is Kit (ex-ConvertKit): it is not in Anthropic's directory; Kit publishes its own MCP URL at kit.com/ai/mcp that you paste in as a custom connector. The directory itself launched July 14, 2025 (Notion, Canva, Stripe, Linear remote; Figma as a desktop extension) with remote connectors paid-only; Anthropic opened connectors to free users on Feb 12, 2026 (free = one custom connector). Notable 2026 additions: Granola MCP (Apr 23, 2026), a nine-connector creative drop (Apr 28, 2026), Gamma (page dated Apr 29, 2026), Microsoft 365 write tools (Jul 7, 2026), Gmail send/reply/forward (Aug 18, 2026, approval on by default). Claude for Small Business (May 13, 2026) is a Cowork toggle bundling QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace and Microsoft 365, not a plan. The Canva and Figma blurbs track the vendors' own docs (Canva MCP: create, edit, search, export; autofill/brand kit are Canva Enterprise-only. Anthropic's Figma page itself says 'audit for design system drift' and FigJam diagrams). Dated security receipt: LayerX's Feb 11, 2026 Google Calendar prompt-injection-to-RCE finding, which Anthropic called out of threat model.

The roster is real; the value question is which three you'll actually use before the permission risk outweighs the novelty.

Confidence

High

Posted by

AIForge (a page)

See the original claim →

The claimReel: "Client wants an 'Amazing Website'… Claude + Framer… The result… I spent only 10 minutes… I didn't write the prompt… Want the exact prompt?" over a dark-blue animated landing page ('Turn balance into momentum', 4.9 / 30k / 15k). Caption: "CLAUDE MYTHOS IS F*CKING INSANE — @viktor just dropped a free 12-minute tutorial on how to build animated, award-winning websites with Claude Fable 5!" 84 likes, 90 comments, 11 shares. Sep 5, 2026.

The reel recycles a three-month-old caption. '@viktor' is Viktor Oddy, a design YouTuber who sells MotionSites.ai prompt packs and the Design Rocket course. His channel has no 12-minute 'Claude Fable 5' tutorial: the two ~12-minute candidates are a Jun 9, 2026 video (12m33s, sponsored by Bolt.new, no Claude in the description) and an Aug 8, 2026 'Fable 5 + Seedance' video (12m40s, Higgsfield affiliate link). His Fable 5 tutorials run 13-60 minutes, and his Sep 2 'Fable 5.1' video is 9m23s. The exact 'Mythos Claude is insane... 12min... award-winning websites with Claude Fable 5' wording first appears on a repost account on X on Jun 12, 2026, the day the US government pulled Fable 5 under export controls. Anthropic's own pages say Fable 5.1 (Sep 1, 2026) and Mythos 5.1 are the same model; Mythos is only for vetted US organizations in the Cyber and Life Sciences verification programs, so no creator is building websites with it. What is real: Claude Design (Apr 17, 2026) and Claude Code with Fable 5/5.1 genuinely produce animated landing pages from a prompt in minutes, and the reel's page (dark-blue, 'Turn balance into momentum', 4.9/30k/15k stats) is the standard MotionSites-style output. 'Award-winning' is marketing: Awwwards scores Design 40 / Usability 30 / Creativity 20 / Content 10 through 18+ jurors over 5 days, and no Claude-generated site is on record winning Site of the Day.

Expect the reel's 'comment for the tutorial' to lead to a DM with MotionSites or Design Rocket links, not a public 12-minute Claude Fable 5 video.

Confidence

High

Posted by

a design-creator page

See the original claim →

The claimNews-card reel: "A 13-year-old programmer is gaining attention online after posting videos of himself solving Codeforces programming challenges in around 45 seconds using Claude. The creator, known as mameeewin… The story is fueling discussion about how younger generations are learning to code alongside AI… an early glimpse into the future of software development and technical education." Clip shows 'A. My First Sorting Problem' solved with a 00:24–00:38 timer. Sep 5, 2026.

The creator is real. 'mameeewin' is a Bangkok middle-schooler with a YouTube channel (136K subscribers as of Sep 5, 2026; joined Mar 18, 2020), a TikTok, a GitHub (30 repos) and a Codeforces account registered Aug 28, 2024 (217 problems solved, zero rated contests). The age is roughly self-consistent: his own YouTube community posts (~mid-2025) say his videos were switched to 'Made for kids' because he was under 13 and that he had just started 7th grade. The speed is real too: his format is 'Codeforces | C++ | Rated 800 | HHKB Type-S | real time | no talking' (e.g. May 28, 2025 upload, 52K views), i.e. a kid typing trivial problems by hand on a mechanical keyboard. What is not real is the Claude part. None of his channel bio, video descriptions, GitHub repos or community posts mention Claude or any AI tool; a May 5, 2026 Instagram summary (9.8K likes) correctly attributes his speed to fast typing. The Claude framing traces to a May 3, 2026 X post ('zero coding experience... he just set up Claude correctly') that linked a generic 'set up Claude' Substack guide which never mentions the boy; a reply noted 'AFAIK no Claude code used in the video.' LinkedIn re-posts (May to June 2026) hardened it into 'using Claude Code,' and this reel, from a San Francisco AI agency's page, is the polished news-card version. The problem on screen, 1971A 'My First Sorting Problem' (Div. 4, rated 800), is 'print the min then the max of two digits.' Any 2026 model answers it instantly; 45 seconds is a human typing time, not an AI benchmark.

IOI 2026 organizers (Euronews, Aug 18, 2026) put it plainly: without the basic foundation you cannot actually use the AI tool, which is exactly what this kid is building by typing it himself.

Confidence

High

Posted by

ASAP Developers (a page)

See the original claim →

The claim"16-minute tutorial on how to build $50,000 cinematic websites (step by step) with Claude Fable 5.1. Save this before you lose it 🔖" Native 16:13 video, no credit. 400K views, 1.3K likes, 120 reposts, 29 replies, 3.5K bookmarks. Sep 4, 2026.

The video is real and the workflow is real. The frames match Nate Herk's (AI Automation Society, ~172K YouTube subs, ~450K Skool members) upload "Fable 5.1 FINALLY Kills AI Website Slop" (youtu.be/FFWtxjvW2ts), posted to his Skool community around Sep 2, 2026 and written up by Geeky Gadgets on Sep 3, 2026 and SolidAITech in Sep 2026. Those write-ups list the same seven steps seen in the frames: brand guidelines, Pain/Person/Promise, scroll, inspiration, components, mobile, verification, with 21st.dev and godly.website named as sources. Nowhere in Herk's title, Skool post, or the write-ups does "$50,000" appear. That number comes from a different creator: Viktor Oddy's Jul 21, 2026 YouTube video "Build $50,000 Websites Using Claude Fable 5 (Full Tutorial)" (~25 min) and his Jul 25 post about a "16-min tutorial." The exact caption "16 minute tutorial on how to build $50,000 cinematic websites (step by step)... Save this before you lose it" has been circulating on X since late July in English and Spanish under other accounts. The poster welded that caption onto Herk's new video, changed "Fable 5" to "Fable 5.1," and gave no visible credit. On price: mid-market cinematic studios (Obys, Locomotive, Basement) do sit in the mid five figures, and Clutch data puts the average agency design project near $56,000, so $50K is a real market tier. A 16-minute tutorial gets you a good-looking landing page, not that tier.

Expect a good-looking, mobile-checked landing page in an afternoon; expect to still need real copy, real photos, and a real reason for visitors to call.

Confidence

Medium

Posted by

@celineodier (Celine Odier)

See the original claim →

The claimPhoto carousel: "Google used Vertex AI Studio and Gemini to show how a working financial analysis assistant can be put together in minutes. The tool can take long earnings reports, pull out structured information, and let users ask questions about the documents directly… For finance teams, that means reports can be turned into searchable, interactive tools much faster than building a custom system from scratch." Slide 1: "The tutorial shows how to use Vertex AI Studio and Gemini to analyze financial reports and build a working AI assistant." Sep 4, 2026.

The source is real: Google Cloud Tech published "How to build a financial analyst assistant with Vertex AI Studio & Gemini in under 10 minutes" on YouTube (video e8hehBcYU7E) on Dec 6, 2025, presented by Colby Hawker and Shubham Saboo, so "minutes" is Google's own framing. It rides the Nov 21, 2025 Google Cloud blog post "Building with Gemini in the newest Vertex AI Studio," which introduced the /prompt, /evaluate and /build agents, Get Code, and login-free Express mode on Gemini 3. What the video actually shows: upload one sample earnings PDF, ask for revenue, COGS and operating expenses, use /prompt to force clean JSON, /evaluate with a rubric that checks 'valid JSON', then /build and Get Code to drop the prompt into an app so a chart library can render it. That is a prompt-and-extraction demo, not a deployed document Q&A system for a finance team. The caption's claims about Studio (prompting, testing, building in one place) match Google's docs. What's missing: the demo runs in Express mode, which Google's docs cap at 90 days without billing, with limited quotas and data sources (Google Drive and web files only); production needs a billed GCP project, IAM, and connectors. Costs (CloudZero, updated Sep 4, 2026): Gemini 3 Flash $0.50/$3.00 per 1M tokens, Gemini 3.1 Pro $2/$12, Search grounding $14 per 1,000 queries after 5,000 free. The FinGround paper (arXiv, Apr 26, 2026) documents that LLMs on filings 'fabricate metrics, invent citations, and miscalculate derived quantities,' with 43% of errors needing arithmetic re-checks. The TikTok caption is a nine-month-old, unlinked paraphrase; a near-identical text ran on Threads Jan 21, 2026.

A ten-minute Google demo proves Vertex AI Studio is quick to prototype in; it does not prove a finance team can replace a custom system without the usual engineering.

Confidence

High

Posted by

@ai.ontiktok

See the original claim →

The claim45-second reel: "There's a way to unlock it… Download Claude Desktop → Help → Troubleshooting → Developer → Configure Third Party Inference → Static API Key → sign up with GitHub [at agentrouter.org] → $125 into your wallet → set it to unlimited → Gateway base URL, bearer auth → Opus 5… running on free credits." 413 likes, 855 comments, 69 shares. Sep 4, 2026.

The clicks in the reel are real. Anthropic shipped 'Claude Desktop on 3P' (Help > Troubleshooting > Developer Mode > Configure Third-Party Inference) as generally available on July 9, 2026, and its docs describe exactly the Gateway/Static API key/Bearer panel shown. But Anthropic built it for regulated enterprises routing through Bedrock, Google Cloud, Microsoft Foundry, or a gateway the company itself operates (LiteLLM, Portkey), pushed by MDM. Its docs say plainly that prompts, files, and tool outputs go only to the configured endpoint and that what happens there 'depends on the provider.' AgentRouter is not on any Anthropic partner list. Its domain was registered Sept 4, 2025 through a Hong Kong registrar with hidden WHOIS, hosted on Alibaba Cloud Singapore; support runs through a QQ group and a Discord (~14k members); sign-up is GitHub-only; guides published Aug 22 and Aug 27, 2026 describe it as an unofficial 'non-profit' relay with no uptime guarantee, no published privacy terms, and daily 'Budget pool quota has been exhausted' 402 errors (GitHub issue, Aug 14, 2026). The reel's own card says Opus 5 is 'Subscription Required,' then claims it runs on free credits. Context: Simon Willison (Jul 26, 2026) and the Cloud Security Alliance (Jul 29, 2026) documented the relay market as pooled/stolen credentials resold at up to 97% off, with prompts harvested for distillation; Anthropic told the Senate on June 10, 2026 that 25,000 fake accounts ran 28.8M Claude exchanges. Real Opus 5 costs $5/$25 per million tokens or Pro ($20/mo) and Max ($100-200/mo). agentrouter.org itself returned 403 to our fetches, so its current pricing and terms could not be verified.

A business owner who follows this reel gets working Opus access for a while, hands every prompt and file to an untraceable middleman, and likely hits daily quota lockouts within weeks.

Confidence

High

Posted by

Aicircl (an AI-training page)

See the original claim →

The claim"SpaceXAI just dropped the best 1-hour course on Grok Bot agents: from 1 prompt to a full 24/7 system. 2:15 Build your first Grok Bot / 6:52 Give every Bot a role / 16:51 Make Bots work 24/7 / 31:50 Run agents in parallel / 52:18 full autonomous system. Free, the best thing on Grok Bot agents I've come across." Native 1:02:12 video watermarked "GrokBotCourse", no source link. 217.1K views, 31 replies, 238 reposts, 1.5K likes, 4.5K bookmarks. Sep 1, 2026.

The product is real. SpaceXAI launched Grok Bot on Aug 11, 2026 as always-on agents with their own cloud computer, a chief-of-staff bot coordinating worker bots, and 24/7 operation. The post is not. Nothing on x.ai/news, docs.x.ai or @SpaceXAI matches a one-hour course with those chapters; the video is a watermarked third-party upload with no link, and the same account posted the identical text and video on Aug 21 before re-posting it Sep 1 with its own article attached. Its feed runs the same template daily, including 'leaked' internal engineering notes that appear as an Anthropic document one day and an xAI document the next with the logo swapped. 'Free' is wrong: Grok Bot requires SuperGrok at $30 a month (Lite at $10 does not include it), with Plus at $100 and Heavy at $300; beta launched Heavy-only with no free tier. SpaceXAI publishes no per-plan Grok Bot quota; the top reply reports burning all tokens in under a day, another says a $200 plan's weekly allotment goes in a day, and Musk issued a one-time bonus token reset on Sep 1 because users hit limits faster than expected. xAI's terms cap liability at $100.

Product real, post is a recycled unsourced re-upload; the honest route is the $30 tier, one sandboxed bot, one task, thirty days.

Confidence

High

Posted by

@Mahaximus_ (Mahax)

See the original claim →

The claimTikTok, Sep 2 2026, 56 seconds (6,741 plays / 276 likes / 18 comments / 280 saves / 89 shares). Caption: "Claude Fable 5.1 + Public databases are an elite combo. (I should be charging for this) #claude #medclinic #leadgen". Title card: "Claude Fable 5.1 + Public Databases = infinite lead glitch". On screen: a Netlify dashboard, "THE MEDICAL CLINIC LEAD MACHINE", showing "5,935 BRAND-NEW CLINICS IN THE TRAY - POURED FROM 105,139 FEDERAL RECORDS", "the 3 weekly files CMS has published since the last monthly cut", "NEWEST WEEK ALONE: 2,048 CLINICS", a banner for "BILLING, CREDENTIALING, MALPRACTICE, STAFFING, EQUIPMENT" vendors, and a sources page stating "COST $0 - no API keys, no scraping, public files". In replies the creator confirms "the npi registry lookup" ("Yup, light work."), that it "all runs in a cloud server with no need for a UI", that it returns phone numbers ("Yup"), and that he runs "5 of 20x plans". A commenter adds "I have this exact set up but I've been using Apollo".

The build is more honest than the hook. On screen it is a real dashboard that reads the three NPPES weekly files (Aug 3-23 2026, 105,139 records), keeps the 5,935 brand-new Type 2 organization NPIs, tags each with the NUCC taxonomy (883 codes), and grades leads on shown math: phone on record, named official is the decision maker, filed in the newest week. It states "COST $0 - no API keys, no scraping, public files", masks names and phones "by policy" and links to the official registry. The timing thesis is sound for billing, credentialing, malpractice, staffing and equipment vendors: a fresh organization NPI marks a clinic's setup window. The data checks out: NPPES is free and public, the API returns addresses[].telephone_number, address_1, city, state, postal_code and address_purpose, and there is no email field anywhere. The framing is the hype. CMS has published these files since the May 30 2007 Data Dissemination Notice; Definitive Healthcare (median $50,533/yr), ZoomInfo ($33,500/yr) and Apollo ($49-$119 per user per month) sit on the same rows. "5 of 20x plans" is $200 x 5 = $1,000/mo to parse three CSVs, and Anthropic's consumer terms bar sharing an account or scripting it outside the API. Nothing on screen shows revenue, and outreach still meets 47 CFR 64.1200 consent rules and CAN-SPAM at up to $53,088 per email.

Free weekly CMS file plus a transparent dashboard is real and useful; "infinite lead glitch", "I should be charging for this" and the five-account Max stack are the hype.

Confidence

High

Posted by

@cfongofficial

See the original claim →

The claimOn-screen "Keep Your Life Private / Peace of mind Every drive / Blocks All Trackers," filmed in a car: a small plug-in device goes into the console port, a phone's location-share map then glitches and whites out. A kraft box with a QR code and a label reading about "DN19014-R ... A1" is shown. Caption "Comment 'privacy' for the link"; the page DMs "Sent! Check your DMs" to each comment. No public link, brand, or price. Posted Aug 12 2026; 1.5K likes, 678 comments, 142 shares, 166K views.

A device that "blocks all trackers" by transmitting is a jammer, and federal law bans marketing, selling, importing, or operating jammers: 47 U.S.C. 302a(b) and 333, with the FCC's enforcement advisory warning of seizure, criminal sanctions, and penalties up to $122,500 per act - illegal even on private property or in your own vehicle. It also fails on its own terms: AirTags and Find My locate via Bluetooth and Apple's billion-device Find My network, not GPS, so a GPS jammer wouldn't stop them, while it would knock out your own navigation and 911. The likely storefront, vanishgear.com, is a three-week-old Shopify dropshipper (ScamAdviser trust score 0/100, hidden WHOIS, internal-only reviews, 10-12 business-day shipping); incognitostick.com 301-redirects to it; its "Incognito Stick" ($80.99) ships "1 Signal Generating Body Unit" plus "2 signal Broadcasting Antennas" claiming a 5-10m disruption radius - the description of a transmitter/jammer. The DN19014 label could not be verified to any specific product (searches returned only unrelated hits - a VW plate, a "Dynasty DN-19014" cutting board).

Skip the illegal jammer; a Faraday pouch and the phones' own unwanted-tracker alerts do the real job legally.

Confidence

High

Posted by

Vanish Gear

See the original claim →

The claimMeme text "Neighbor: We won't lower our music / Me:" cuts to two black SMA antennas in hand and a bare USB-C-powered PCB dev board (an ESP32-class module with a shielded can and pin headers), red LED lit. Caption "Comment 'silence' for the link"; the page replies "Sent! Check your DMs." No brand, model, or price shown. Posted Aug 20 2026; 883 likes, 570 comments, 49 shares, 73K views.

The hardware shown - a dual-antenna ESP32-class 2.4GHz board - matches the open-source "ESP32-BlueJammer" genre (github, ~7.9K stars) whose own README states plainly "Jamming is ILLEGAL! Educational purposes only!" Such a board disrupts the whole 2.4GHz band: Bluetooth, BLE, and Wi-Fi alike, which is exactly what would silence a neighbor's speaker - and exactly what federal law forbids. The FCC's enforcement advisory (fetched) says it is illegal to operate a jammer anywhere, including in a residence or on private property, illegal to sell or advertise even a single unit, and that importing one makes you the importer of an illegal device; 47 U.S.C. 302a/333 back this, with penalties up to $122,500 per act plus criminal sanctions. There's also a technical bait-and-switch in this product genre: a cheap ESP32 "deauther" only sends Wi-Fi deauthentication frames and does not affect most Bluetooth speakers; a board that truly jams Bluetooth is a broadband 2.4GHz noise transmitter - the illegal part - and it simultaneously clobbers your own Wi-Fi and can block 911. As with the other reel, there is no link, brand, or price - only a DM funnel.

Handle the noise through ordinances and the landlord; the advertised board is an illegal jammer, not a fix.

Confidence

High

Posted by

Vanish Gear

See the original claim →

The claimCaption: "Millionares are born with AI everyday #krea #aiautomation". On-screen: "This is our fun little side hustle" / "After buying all these scratchers we made $800" (borrowed lottery-ticket clip), "DON'T DO THAT / DO THIS INSTEAD", "GO TO CLAUDE", "CUSTOM CONNECTOR", "GO TO KREA.AI/MCP", paste https://api.krea.ai/mcp, paste a GTA 6 trailer YouTube link, prompt "Create me a 30 day content plan to make $10,000/month making GTA 6 content", "USE KREA TOOLS TO MAKE GTA VIDEOS ... FOR YOUR YOUTUBE", "COMMENT BELOW GTA". 2.9K likes, 1.7K comments, 361 shares; posted Sep 1 2026.

The only verifiable part is the connector. Krea's MCP page (published Jul 20 2026) lists https://api.krea.ai/mcp with Krea-account sign-in and no API key, and Anthropic's help center allows one custom connector on Claude's Free plan. Everything after that collapses. Grand Theft Auto VI has not shipped; Rockstar's official page gives a Nov 19 2026 launch, so 'GTA 6 content' in September 2026 means trailer reaction or AI-fabricated GTA-style clips built from Rockstar's trailer as a reference. Rockstar's posting policy allows fan use only when non-commercial, defined as not making money through the footage, and its terms bar using its material as source for Generative AI tools. YouTube's Partner Program requires 1,000 subscribers plus 4,000 watch hours in 12 months or 10 million Shorts views before any ad revenue, with roughly a month of review, and its inauthentic-content policy (effective Jul 15 2025) excludes AI content made with generic templates that gives the impression of mass production. A 30-day plan from Claude is a document, not a projection; the reel shows no channel, no analytics and no dollar earned. The FTC banned Ecommerce Empire Builders in 2025 for promising '$10,000 each month'. The opening scratch-ticket clip is borrowed and unrelated, and 'comment GTA' is comment bait under Meta's own definition.

Connector real, income route blocked by GTA VI's unreleased status, Rockstar's non-commercial rule and YouTube's monetization thresholds and AI policy; redirect to owned-IP video.

Confidence

High

Posted by

Jamen Creates

See the original claim →

The claimCaption: "shoutout zuckerberg for this one doe #krea #hustling #meta #sidehustle". On-screen: "Day 2 of trying RANDOM side hustles" (randomizer lands on META ADS), "15 minutes for research", "click on Connectors", "copy this link" (Meta Ads MCP Server doc, mcp.facebook.com/ads), "now use the Krea MCP", "enter your product", a 20-image brand-campaign prompt for a CPG sunscreen, "fund your campaign", "comment 'Krea' ... you the link". 5.2K likes, 332 comments, 668 shares; posted Sep 2 2026.

The two tools shown are real and cost nothing to wire up. Meta's developer docs (Ads MCP Server, updated Jul 14 2026) describe a Meta-hosted MCP server at https://mcp.facebook.com/ads that creates and edits campaigns, ad sets and ads, pulls reporting, manages catalogs and runs A/B tests; it authenticates through Facebook Login for Business with ads_management and related permissions, and the endpoint answered a live JSON-RPC probe on Sep 2. Krea's MCP page (published Jul 20 2026) gives https://api.krea.ai/mcp, no API key, sign in with a Krea account, works with Claude, Codex, Cursor. Anthropic's help center allows one custom connector on the Free plan. What the reel omits is everything that costs or makes money. Krea's free tier is a daily compute allowance and its FAQ lists the commercial license only on paid plans (Basic $9/mo for 5,000 units, Pro $35 for 20,000, Max from $70), so ad images made on Free are not licensed for ads. The campaign is for a sunscreen that does not exist, so no product, margin or revenue is shown; Shopify's Nov 2025 benchmarks put Meta ads at about $0.87 per click and $18.75 per lead. No Krea affiliate program was found (krea.ai/affiliate 404), so the poster's payoff is the comment-to-DM funnel, which Meta's help page defines as comment baiting and says receives less distribution.

Tools verified real and free to connect; the business case is empty, so the value is a Meta-ads reporting and creative-testing workflow for owners who already advertise.

Confidence

High

Posted by

Richky Lim

See the original claim →

The claimOn-screen "Claude vient de tuer les web designers" (Claude just killed web designers) over the dark animated "ORYZO - Made for mugs. Built for tables." site. Steps: install design skills "emil-kowalski" (motion/animations) and "impeccable" (UI/UX playbook), add a skill that draws from real sites, connect Figma MCP + Playwright so Claude "designs, builds, tests automatically" - then "describe your site... a professional site in a few minutes." Caption "Follow @jb.roy_ pour implementer l'IA"; comment DESIGN, must be subscribed. Posted Aug 14 2026; 1.2K likes, 881 comments, 100 shares.

Every named tool is real and mostly free. Emil Kowalski's skills repo (github.com/emilkowalski/skills, ~34.9K stars, MIT; install npx skills@latest add emilkowalski/skills) ships animate/motion/apple-design skills; "impeccable" is Paul Bakaus's open-source design skill (github, ~65K stars, Apache-2.0, npx impeccable install). Both are third-party, not Anthropic. Figma's Dev Mode MCP (remote server free in beta, supports Claude Code) and Microsoft's Playwright MCP (Apache-2.0, free, ~36.8K stars, supports Claude Code) exist. Claude Code does generate full front-ends, and Anthropic launched Claude Design Apr 17 2026, included with Pro ($20/mo)/Max (from $100)/Team/Enterprise. The overclaim is the demo: the ORYZO site is a self-initiated satirical concept by Lusion, a Bristol WebGL studio - oryzo.ai's own footer reads that it is "a fictional creative project by Lusion," it won Awwwards Site of the Day (Apr 13 2026, score 7.86) and was built with GSAP, Three.js and WebGL by professionals. And designers aren't dead: BLS projects web developers/digital designers to grow 5% 2025-2035 (faster than average, 220,100 jobs, median ~$92,650), noting AI "may soften" growth, not end it.

Use the free tools with realistic expectations and keep human design judgment in the loop for anything client-facing.

Confidence

High

Posted by

JB Roy

See the original claim →

The claimFacebook reel, Sep 2 2026 (651 likes / 269 comments / 108 shares). Caption: "Simple swaps in your videos! Comment \"AI\" to try it yourself! higgsfield_ai. Gen jutsu runs on Seedance 2.5 and lets you replace anyone or anywhere in an existing video, so you can recast a scene or move it to a completely different setting without reshooting. It gives creators total control over casting and location after the footage already exists, which is huge for filmmakers, advertisers, and content creators working at speed. #HiggsfieldAI #GenJutsu #Seedance #AIVideoEditing #CreativeAI". Poster frame: "Higgsfield GENJUTSU — Infinite Content Era", split screen of a woman in a kitchen swapped for a bearded man in a cap. Comments are a flood of "AI" plus "Am freaking buying higgfield".

The feature exists. Higgsfield's Genjutsu page is live, the homepage bills it as the "New Model", and the launch post is dated Aug 31 2026, two days before this reel. It does what the caption says in two modes, Motion Transfer and Object Swap. Seedance 2.5 is also real; ByteDance Seed announced it Jul 31 2026 and Higgsfield hosts it. What the reel adds on its own authority is "runs on Seedance 2.5": neither the Genjutsu page nor the launch post names an underlying model, and the specs do not match (Genjutsu tops out at 1080p and 30-40 reference images; Seedance 2.5 on Higgsfield is native 4K with 50). Higgsfield's blog puts a 15-second 1080p Genjutsu run at 144 credits, about $7.20, on plans from $49/mo for 1,000 credits, no free tier for this model, and concedes "real cost per clip" is generation cost times regenerations. "Comment AI" is the affiliate program: up to 25% of your subscription for up to 12 months, with a 50%-off offer that expires three hours after you click, and no disclosure on the reel. The privacy policy says uploads, prompts and outputs train Higgsfield's models and cannot be removed once trained; some features take a faceprint. The Terms of Use page could not be retrieved, so likeness and watermark rules stay unverified.

Real feature, real model, unverified "built on Seedance 2.5" claim, and an undisclosed affiliate funnel with the price and training-data terms left out.

Confidence

High

Posted by

Rourke Heath

See the original claim →

The claim"editors, are we cooked?" then a screen recording captioned "1. upload all your raw footage", "2. tell stanley studio to edit whatever you want" (prompt: "Make it a punchy 9:16 reel, Add background music"; model selector "GPT-5.6 Sol Low"), "3. let it cook" (agent log at studio.getstanley.ai: Add Text, Add Broll Shot, Add Zoom Region, Add Cut Sfx, "Syncing the beats"), "4. get a fully edited video in minutes", "comment 'EDITOR' for the link". Caption "#editing #contentcreator #ai #startup". Posted Aug 27 2026; 103K views, 1K likes, 82 comments, 108 shares.

The product is real. Stanley Studio (studio.getstanley.ai) is a browser video editor with a chat agent, built by Stan, the creator-storefront company (getstanley.ai reads "Part of Stan Store"; the Product Hunt listing names makers Daniel Park, Justin Badua and Oliver Chen as "the team behind Stan", 202 upvotes, launched roughly July 2026). Its pricing page says Free: $0, unlimited projects and exports, "$50 in AI credits to start", no card; Pro: $49/month for "$1,000 in AI credits every month". So the comment war ("Not free" vs "This free") has an answer the reel never gives: free until the $50 of credits run out; the site never says what one edit costs. "In minutes" matches the vendor's own copy and is plausible for cuts, captions and music; whether the cut is good is never shown, since the finished video is not played. The affiliation inference holds: the poster's Slack tab reads "Oliver Chen (DM)", the name of a listed maker, and a Notion tab reads "Stanley Studio - Fiona". Stanley also runs a Partner Program giving every Stan product free ("$249/month value") for a bio link and 10 referrals a month. The caption discloses nothing. "Editors are cooked" fails the labor data: BLS projects film and video editors and camera operators up 3 percent 2025 to 2035, 72,000 jobs, median $75,100.

Real free-to-start tool with an undisclosed $49 tier and likely insider posting; try it free, verify the output yourself.

Confidence

High

Posted by

momostudio

See the original claim →

The claim"Top 5 tools for your AI Agents": "if you wanna give your AI agent its own email, use AgentMail"; "let your agent log into any profile, platform or software and only verify once, use Composio"; "give your agents their own computer and watch them work, use Orgo cloud"; "access clean web data, use Firecrawl"; "safely execute code, use E2B"; "if you're looking for more of the best AI use cases for absolutely free, give me a follow." 30-second reel, July 12, 2026. 260K views, 4.5K likes, 109 comments, 481 shares.

Every tool on screen is real, current developer infrastructure and each is described in roughly its own words. AgentMail provisions API inboxes for agents: free tier 3 inboxes and 3,000 emails a month, $20 a month for custom domains. Composio handles OAuth and API keys for 1,000-plus apps so an agent authorizes once: 100,000 free tool calls a month, Pro $29. Orgo sells cloud desktops agents can drive: no free tier, prices only visible after login, usage billed while a machine runs. Firecrawl turns pages into LLM-ready markdown: 1,000 free credits a month, Standard $83 for 100,000, and it respects robots.txt by default with overrides reserved for enterprise. E2B runs sandboxed code per second of runtime: $100 one-time credit, one-hour sessions on the free tier, Pro $150 with 24-hour sessions. The reel makes no income claim, does not call the tools free, and its only ask is a follow. It went real rather than partly because the omissions are the ordinary limits of a thirty-second list, not a hidden catch: what it skips is domain authentication and inbox warm-up for agent email, per-second compute bills, Orgo's undisclosed pricing, and the fact that an agent holding email, logins, a desktop and a code sandbox needs approval gates before it touches anything that costs money.

Real tools, honest ask; the work is the plumbing, so start receive-only on the free tiers and add layers one at a time.

Confidence

High

Posted by

Hegetfunds

See the original claim →

The claimOn-screen "This is the fastest way to rank higher on Google," caption "You're welcome!" The mechanism, drawn out in the comments, is to run a Google search operator - site:yourdomain.com "keyword" - to find your own pages that already contain a keyword, then add internal links from those pages to the page you want to rank. Commenters: "It's all about the colon," "It's not the pages that rank for the keyword, it's the pages that contain the word," "He is talking about Internal Linking." Posted Aug 26 2026; 571 likes, 18 comments, 45 shares, 97K views.

The tactic is legitimate and straight out of Google's own guidance. Google Search Central says every important page should be reachable by at least one internal link with descriptive anchor text, and John Mueller called internal linking "super critical for SEO... one of the biggest things you can do on a website" (widely reported, Mar 2022). Two caveats make it "partly." First, "fastest way to rank higher" oversells timing: Google's own SEO documentation says changes can take "a few hours" to "several months" and advises waiting a few weeks to judge impact - internal links help crawling and relevance but aren't an instant lever. Second, the site: operator is a discovery trick, not a source of truth: Google explicitly documents that site: "doesn't necessarily return all the URLs that are indexed" and points users to Search Console's URL Inspection instead. HeyTony is a real, verifiable agency (heytony.ca) in Hamilton, Ontario, founded by Matt Diamante, roughly 20 staff, author of the SEO book "Get Found"; the reel is genuine advice that also feeds a funnel toward the paid "HeyTony Insiders" community at $97/month.

Add the internal links and verify indexing in Search Console; it's a real habit, just not an instant ranking switch.

Confidence

High

Posted by

HeyTony

See the original claim →

The claimBlack-slide video: "7 GitHub Repos = Master AI Engineering" listing 1/ Awesome Artificial Intelligence, 2/ Awesome LLM Apps, 3/ 100 Days of ML Code, 4/ System Prompts and AI Tools, 5/ AI Agents for Beginners, 6/ Microsoft Gen AI for Beginners, 7/ Learn Agentic AI; "Comment 'AI' for all Repo Links". Caption: "Comment "AI" for all links. Master AI Engineering using GitHub repos. Save it & follow for more." Posted July 24 2026; 191K views, ~2K likes, 529 comments, 322 shares.

All seven repos exist and are public, so no comment is needed. GitHub API counts on Sep 2 2026: owainlewis/awesome-artificial-intelligence 16,309 stars (a link list started 2015); Shubhamsaboo/awesome-llm-apps 135,806 (100+ runnable open-source agent and RAG apps, Apache-2.0); Avik-Jain/100-Days-Of-ML-Code 51,702 (last push Dec 2023, dormant); x1xhlol/system-prompts-and-models-of-ai-tools 143,304 (the slide's title is wrong; this is a dump of extracted proprietary system prompts from Cursor, Devin, Claude Code, Lovable and others, with a "LeaksLab Discord" badge and crypto donation addresses); microsoft/ai-agents-for-beginners 73,784 (18 lessons); panaversity/learn-agentic-ai 4,351 (course material for a paid Panaversity certification, last push Oct 2025). Slide 6 says "Microsoft Gen AI for Beginners" (microsoft/generative-ai-for-beginners, 119,041 stars, 21 lessons) but shows the sketchnote from microsoft/AI-For-Beginners (67,980 stars, 24 lessons), a different course. The verb is the lie. Two are link lists, three are beginner curricula, one is a leaks archive, and none includes a project you ship, feedback, or a portfolio, which is what employers mean by AI engineering. Meta's own help center defines comment baiting as "asking people to comment with specific answers: words, numbers, phrases or emojis" and says such posts "will receive less distribution".

Repos real, verb false; do one Microsoft course and ship one small app instead of collecting links.

Confidence

High

Posted by

codeslearningoct24

See the original claim →

The claimA reel reacts to a purported @elonmusk tweet (shown dated Aug 3, 2026, with 2.4M views): “Running Kimi K3 locally is easy. You just need: 1.5 TB of GPU memory, roughly 8 H100s, a small power substation.” Caption: “Can you really run Kimi K3 locally? 😅 Apparently, all you need is 1.5TB of GPU memory, 8 NVIDIA H100 GPUs, and a small power station.”

Grade the artifact, not the vibe — and the artifact is a fake receipt. The joke tweet is real, but it was posted by a small account (@shiri_shh), verbatim, during the week Kimi K3’s weights actually dropped in late July 2026. The version in this reel has been re-skinned with Musk’s handle, a wrong date, and an invented 2.4M view count; no such Musk tweet exists in any index. Musk’s entire documented K3 engagement is a one-word “Impressive” reply under a benchmark post and a boast that xAI’s next model would surpass it. The delicious part: the fabricated tweet lowballs the truth. Kimi K3 is real — Moonshot AI’s 2.8-trillion-parameter open-weights MoE — and 8×H100 is only 640GB of memory, less than half of what even a 4-bit quantization needs (~1.4TB). A practical deployment is roughly forty H100-class GPUs at an estimated $60K–$115K a month. So the reel’s own numbers contradict each other by a factor of ~2.4, under a name that never wrote them.

Kimi K3 and the joke are real; the Musk attribution, date, and view count are fabricated — and the fake tweet actually understates the hardware by half.

Confidence

High

Posted by

Sanjay Behera

See the original claim →

The claim“Fable 5 and my employees working hard” / “They never sleep” — a phone-farm rack of dozens of smartphones presented as AI ‘employees,’ on an account whose stated theme is phone farming (pinned videos: ‘Just post more,’ a Claude-logo ‘I did it’). 9.3K likes, 1.8K saves.

Two true ingredients, one false dish. Claude Fable 5 is real — Anthropic’s 2026 flagship, genuinely state-of-the-art at long-running agent work — and phone farms are real: racks of devices run 24/7 to fake views, likes, follows, and app-store rankings. But a phone rack isn’t ‘AI employees.’ It’s an engagement-manufacturing rig, and nothing in Anthropic’s product drives racks of physical phones — platform manipulation and spam are explicitly against Anthropic’s usage policies. TikTok’s community guidelines ban exactly this, down to banning instructions for doing it, with penalties from stripped metrics to permanent bans and monetization blocks; mass ban sweeps of device farms are routine. Even the funded, ‘legit’ version of this business — the a16z-backed phone-farm startup Doublespeed, whose cofounder called Claude Code ‘truly our third cofounder’ — is described by reporters as flooding social media with AI slop. The reel borrows a frontier-lab model name to brand a spam rig.

Real model name, real phone racks, fake premise — engagement farms are banned by every major platform and Claude doesn’t run them.

Confidence

High

Posted by

@ifmowgliwasrich (Mowgli)

See the original claim →

The claim“100 Claude Code Commands part two” — a slash-commands rundown for Anthropic’s Claude Code CLI, with a live demo of a pizza-ranking map app (‘Your restaurant 4.4’ among local competitors). 14.3K likes, 5.3K comments, 1.5K shares; the companion cheat sheet is comment-gated (‘Comment Claude’).

The biggest post of this batch is also one of the most defensible — mostly. Claude Code is real, the slash commands are documented by Anthropic, and the pizza-map demo is exactly the kind of small app the tool genuinely builds well. The stretch is the number: the official commands reference lists roughly 65+ built-ins (/help, /login, /doctor, /debug, /usage, /compact, /model, /mcp, /permissions and the rest). ‘100’ only gets there by padding with aliases (/cost = /usage, /quit = /exit) and custom commands — which are unlimited by design, because any file you drop in .claude/commands/ becomes one. That’s a real feature, but it means the ‘100’ says nothing about the product; you could as honestly claim 1,000. And the cheat sheet behind the comment gate is a repackaging of a free public docs page.

Real tool, real demo, padded count — ~65+ documented built-ins, and the gated cheat sheet is a free docs page in disguise.

Confidence

High

Posted by

Jonathan Acuna

See the original claim →

The claim“These three websites gave me 500+ AI API keys for free! I found a way — Get 500+ AI API keys.” (Comment-gated for the list; 87 comments.)

No legitimate service hands out 500 API keys. The number almost certainly comes from Puter.js marketing — ‘One API for 500+ AI models’ — which is 500 MODELS behind ONE key on a user-pays gateway, a completely different thing. The real free landscape is modest and single-key: OpenRouter’s free models (~50 requests/day base), Groq’s free tier, Google AI Studio’s Gemini tier (~500 free requests/day) — free TIERS you sign up for yourself, with daily caps measured in dozens to hundreds of requests. The dark side of this genre is real: GitHub lists that offer working ‘keys’ you didn’t create are trafficking in scraped or stolen credentials — the practice security vendors call LLMjacking — and there’s a documented case of one stolen cloud key racking up $82,314 in charges in 24 hours. Twelve thousand-plus real keys have been found exposed in public repos. Either the video means free tiers (three keys, not 500) or it means keys that belong to someone else.

Free tiers are real and small; ‘500+ free keys’ is either a garbled gateway pitch or stolen credentials. Nothing in between.

Confidence

High

Posted by

Learninwithak

See the original claim →

The claim“I charge my clients $14,630 per video edit…” thanks to two ‘AutoEdit’ AI plugins for Premiere Pro. “Comment ‘SUITE’ to get these 2x AutoEdit AI plugins.”

We went looking for $14,630 and found nothing: no pricing guide, no rate card, no invoice standard carries it — and neither do the vendor’s own product and marketing pages, which make zero income claims. Real freelance video-edit project rates run $300–$1,500 typical and $2,500–$7,500 for large multi-deliverable campaign work; $14,630 for one edit is two to ten times the top of the professional range. The oddly-specific figure (not $15K — $14,630) is a known ad-copy trick: fake precision reads like a real invoice. The plugins behind the comment gate are real — Music Video Mode at $179.99 one-time and Creator Mode from $9.99/mo — both sold by the vendor as rough-cut tools you refine by hand. The products are ordinary and affordable, which is exactly why the ad needed an imaginary invoice.

Real ~$180-and-under plugins, invented invoice — the $14,630 exists only in the ad. See #239, #238, #227.

Confidence

High

Posted by

Eleven Percent LLC

See the original claim →

The claim“POV: you charge clients $10K per video edit…” — an AI Premiere Pro plugin (AutoEdit Music Video Mode) edits music videos by itself while the creator works out. “Comment ‘Autoedit’ to get the link.”

The plugin is real — and its own product page kills the premise. AutoEdit Music Video Mode ($179.99 one-time) does BPM detection and beat-synced cuts, flags shaky B-roll, and ‘creates the first rough cut’; the vendor says in plain text that it is ‘not a full video editor’ and saves ‘the first 3–4 hours of setup work,’ after which you refine manually. Nothing edits a finished music video while you do pushups. And the invoice math is theater: professional video editing typically bills $1,000–$5,000 per project ($20–$150+/hr), while ENTIRE independent music video productions — shoot, crew, gear, and the edit — run $10,000–$30,000. ‘$10K per edit’ quotes a full production budget as if it were the editing line item. The comment gate is the point: aspiration-bait pricing drives DMs to a ~$180 plugin.

Real $179.99 rough-cut plugin; the unattended editing and the $10K invoice are the ad. See #240, #238, #227.

Confidence

High

Posted by

Eleven Percent LLC

See the original claim →

The claim“POV: you got tired so you let claude AI edit your content in under 2mins” — “Comment ‘claude’ and I’ll dm you AutoEdit Content Creator Mode plugin for free.”

Fourth time this plugin has crossed the audit (#061, #214, #227) and the read hasn’t moved: AutoEdit Creator Mode is real and genuinely Claude-powered — a third-party Premiere plugin that transcribes your footage, sends the transcript to Claude, and assembles a rough cut. There is no official Anthropic Premiere plugin; this is a vendor calling Claude’s API, which is fine and worth saying precisely. The lie is the same one as always: ‘free.’ The vendor’s pricing page says the $0 tier has ‘no access to Creator Mode AI editing’ — free gets four utilities (paste image, silence cutter, sound effects, motion graphics); the Claude editing shown starts at $9.99/mo with video and minute caps. And ‘under 2 mins’ oversells a tool whose own vendor describes rough cuts and first drafts you then refine. Real product, honest $10 on-ramp, dishonest hook — every single time it’s advertised.

Genuinely Claude-powered rough-cut plugin; the ‘free’ and the 2-minute finish are the ad. Fourth sighting — see #061, #214, #227.

Confidence

High

Posted by

Eleven Percent LLC

See the original claim →

The claim“In a 50-minute podcast, a SpaceXAI engineer (ex-Cursor) showed how to build a team of agents that will work for you 24/7” — watermarked ‘GrokBot team workshops.’ 37.5K likes, 36.1K saves.

The core is true and traceable: the person is Lauren Tan — genuinely ex-Cursor (and Meta and Netflix), now an engineer at SpaceXAI working on GrokBot — and the ‘podcast’ is a real recorded ~50-minute team workshop in which she shows how she runs 10–20 agents, with about 90% of her routine work automated through a ‘Chief of Staff’ agent, including agents fixing bugs while she sleeps. Clips of it pulled millions of views, and the source material is free. What the TikTok garbles: it’s a workshop, not a podcast; and ‘agents that will work for YOU 24/7’ quietly swaps a frontier-lab senior engineer’s hand-built setup for a promise to the viewer. The watermark account is a repackaging channel riding the moment — a cottage industry of paid agent-team masterclasses sprang up around a talk that costs nothing to watch.

Real ex-Cursor engineer, real free workshop, real 10–20-agent setup — repackaged by a clip account into a 24/7 promise the talk never makes to viewers.

Confidence

High

Posted by

@sarutalksai

See the original claim →

The claim“Go to grader.owner.com to see how you can improve your restaurant’s online presence” — a video promoting Owner.com’s 90-second restaurant-website grader.

Everything named is real. Owner.com is a $1B-valuation restaurant-marketing SaaS ($120M Series C in 2025, ~$189M raised, est. $81M ARR), and grader.owner.com works as advertised — we ran it live: enter a restaurant, it scans SEO, site issues, and competitors into an ‘online health score.’ It is unambiguously lead generation: the free scan is the top of a funnel that lands at $249–$499/month plus roughly $1,000 implementation. The asterisk is disclosure. Lemonis is a genuine Owner.com investor — he’s said so himself on X (‘I invested in this business two years ago’) and participated in the 2021 seed — but this Facebook post carries no #ad and no investor note, so viewers see a neutral tip rather than a portfolio plug. Real tool, real value in the diagnostic, know whose funnel you’re standing in.

Real $1B company, real working grader, real (elsewhere-disclosed) investor — a legit free diagnostic that doubles as the top of a $249–$499/mo funnel.

Confidence

High

Posted by

Marcus Lemonis

See the original claim →

The claimReposts a Bilawal Sidhu tweet — “Making 3d motion graphics with coding agents is absurdly fun. Breaking down geospatial tech for my next video” — over a green wireframe satellite-terminal demo stamped ‘SYNTHETIC DEMONSTRATION — NOT AN OPERATIONAL SYSTEM / PANOPTES CORE.’ Funnel: ‘Comment 3d for my satellite deep dive.’

A viral AI reel where the person, the tool, and the output all check out. Bilawal Sidhu is exactly who the reel implies: ex-Google Senior PM for 3D Maps and spatial computing, TED AI host, and he verifiably builds this exact thing — his agent-built geospatial ‘WorldView’ command-center went viral with ~4M views in April 2026, and he open-sourced a browser spy-satellite simulator built with multiple AI agents, complete with the same intel-terminal aesthetic. Coding agents genuinely script three.js/WebGL scenes and Blender; his shipped repos are the proof case. We couldn’t pull the verbatim tweet from search indexes (X’s coverage is spotty), so one hedge stands — but every element matches his documented public work, and the ‘SYNTHETIC DEMONSTRATION’ stamp is an honesty marker you almost never see in this genre. The only funnel is a comment-bait ‘3d’ trigger on the reposting account.

Real person, real open-sourced agent-built 3D work, honest self-labeling — the batch’s proof that the genuine article doesn’t need a fake invoice.

Confidence

Medium

Posted by

Billy3d

See the original claim →

The claim“I charged $7K per project using these exact Figma files. Now all 39 are yours, high-end animated prototypes, fully editable, running on Figma’s free plan. Instant access.”

We traced the funnel to its landing page and ‘instant access’ means ‘instant checkout’: the 39-prototype pack is a paid product — $39 list, $19.90 ‘launch price,’ delivered by email after card payment. The post’s gift framing fails on contact, and the page has since drifted to ‘42 files’ with ‘2 new templates every week’ — a content treadmill, not a vault of client deliverables. The $7K-per-project claim has zero corroboration: the sales page itself shows no clients, no testimonials, no portfolio, and makes no income claims — the number exists only in the ad. ‘Running on Figma’s free plan’ is trivially true of any shared Figma file. And the product category’s market price hovers near zero: Figma Community and freebie aggregators carry thousands of dark futuristic UI kits for nothing. Selling $20 template packs by implying they’re the secret behind $7K invoices is the design-world MRR pitch.

A $19.90 tripwire dressed as a gift, with an income claim that exists nowhere outside the ad.

Confidence

High

Posted by

Grigoletto.ia

See the original claim →

The claim“If you were born between 1959-1997” — “If you’ve been thinking, ‘I could sell digital products... But what could I sell and how do I sell it?’” (comment-gated funnel).

Start with the hook: 1959–1997 spans roughly ages 28–67 — essentially every working-age adult on Facebook. It excludes no one; it exists to make the viewer feel selected. No credible source ties birth years to digital-product income. The script underneath (‘what could I sell and how do I sell it?’) is the signature opener of the Master Resell Rights funnel economy — the wave of $400–$500 courses where, as reviewers document, roughly 90% of buyers end up reselling the course itself because it’s the only ‘product’ they have. We couldn’t trace this specific page’s back-end (no footprint outside Facebook), so we grade the archetype it matches — in a niche where the FTC has recently banned multiple ‘make money online’ operators for life over deceptive income claims and processors have squeezed MRR sellers over income-claim violations. When the targeting is fake, assume the opportunity is too.

Fake-precision targeting (everyone born in a 38-year window) wrapped around the standard MRR funnel script; the specific back-end has no verifiable footprint.

Confidence

Medium

Posted by

Wealth with Jennifer

See the original claim →

The claim“Clippers make more money than 9-5 workers” — over an Adin Ross Kick-stream clip (‘ADIN ROSS DID WHAT?!?’). Ross’s real quote: ‘Why do you guys work 9-5 jobs when you can make $50 per 100,000 views? My clippers make more money than 9-5 workers.’

The clipping economy is real — that’s why this isn’t a full BS. Streamers and platforms genuinely pay bounties for viral clips, and the payouts are documented: campaign rates run roughly $0.20–$6 per 1,000 verified views, averaging around a dollar. Now run Ross’s own advertised number: $50 per 100K views is $0.50 per thousand — to out-earn even a $40K salary you’d need about 6.7 million verified views a month, every month. The market’s median tells the real story: full-time solo clippers land around $300–$2,500 a month, and the people genuinely beating a 9-5 are a small managed tier on agency retainers. Financial reporting on the trend adds two caveats the reel skips: typical side-hustle earnings are modest (~$200/month), and clipping farms have amplified false accusations that couldn’t be contained. A real micro-gig with lottery-ticket outliers, sold as a salary replacement.

Real economy, real quote, wrong math for the median — $0.50/1K views needs ~6.7M monthly views to beat a $40K job.

Confidence

High

Posted by

Easymoneymike

See the original claim →

The claim“If I lost everything, I would start a secret Instagram account with Claude. Here are 7 prompts to build it from zero.”

The prompt ladder is real and Claude handles it well — niche research, pain points, 60-second scripts, hooks, b-roll sourcing, monetization ideas, a 30-day calendar. Near-identical reels run across the faceless-IG genre (one on a 9M-follower account uses the same seven steps), and prompt-pack roundups for exactly this exist. That half stands. The fantasy half is everything the reel skips: Claude has no native Instagram posting; the auto-engagement tools that fake it violate Instagram’s Terms of Use and trigger enforcement; and a ‘from zero’ account earns nothing until a human films, posts, engages, and grinds through the months when nobody watches. There’s no evidence this specific creator (62 likes) built any income this way — the reel is a repackaged prompt list, and the typical funnel behind it is a comment-to-DM automation delivering a lead magnet or a paid prompt pack.

Real prompts, real planning value; the posting, engagement, and income are all still human work the reel waves away.

Confidence

High

Posted by

Digitalwithdp

See the original claim →

The claim“This boy added 200+ free AI tools, coding agents, automation, SEO all in one place” (comment-gated for the link; 1.1K comments).

The claim isn’t fake — it’s worthless as a secret. AI tool directories with far more than 200 entries exist by the dozen: the big aggregators claim 2,000 to 30,000+ tools each, publicly, no comment required. So ‘one site with 200+ tools’ is trivially true and not special; the creator compiled links, he didn’t build tools. Two catches ride along. First, ‘free’: directory listings are overwhelmingly freemium — free tier plus paid upsell — and reviewers warn the listings run stale on pricing. Second, the industry’s own conflict: major directories have moved to pay-to-be-featured models where rankings reflect submission fees and affiliate value, not quality. The comment gate is the actual product — it harvests your DM for a funnel, delivering a link you could have googled in ten seconds.

True but trivial — public directories dwarf it, ‘free’ means freemium, and the comment gate is the real product.

Confidence

High

Posted by

Learninwithak

See the original claim →

The claim“Master AI Coding with these 6 Github Repos” — whiteboard list: Caveman, Superpowers, RAGFlow, Andrej-Karpathy skills, Awesome-Claude plugins (+1).

Every repo on the whiteboard exists, and several are genuinely excellent: Superpowers is Jesse Vincent’s MIT-licensed agentic-skills framework for Claude Code (on the official plugin marketplace), RAGFlow is a serious Apache-2.0 RAG engine and one of GitHub’s fastest-growing projects, and Karpathy’s teaching material — nn-zero-to-hero, nanoGPT and its successor nanochat — is arguably the best free AI education anywhere. But the ‘expert’ framing is the con, on two counts. Two of the six are literally lists of links to other repos (the awesome-* collections) — bookmarks, not training. And Caveman repeats a finding from our August audit: its advertised ~65% token savings held only on chat-style tasks, with independent tests measuring 8.5–14% on real agentic coding runs. Starring repos is bookmarking. Building through one of them is education.

Six real repos, two of them just link lists, one with a debunked stat — one repo built-through beats six starred.

Confidence

High

Posted by

Kushal Vijay

See the original claim →

The claim“Ranking AI Website Builders OR Video Generators” — an S-to-D tier list with Lovable placed first. Caption: ‘save these tools for later bro trust me.’

There’s almost nothing to debunk here, which earns it a real — real-but-empty. The products in the ranking exist, and the one shown going on the board first is legitimately enormous: Lovable, the Stockholm AI site builder, raised $400M at a $13.3B valuation in August 2026 with ARR approaching $600M, after the fastest $1M-to-$100M ARR run in software history. A tier list is opinion; rankings in this genre are usually engagement bait rather than hands-on testing, and this one’s near-zero engagement suggests even the bait didn’t bite. The only honest caution: nobody’s S-tier placement should choose your stack.

Real products, opinion ranking — the shown leader really is a $13.3B company; the list itself carries no claim worth grading.

Confidence

Medium

Posted by

Aidenlovesai

See the original claim →

The claim“POV: Claude x video editing = $12k/month. 1. I use the free Claude plugin to edit shortform content. 2. It adds captions and edits viral shortform + it’s free.” Caption: “this Claude AI video editing method is INSANE — Comment CLAUDE for the FREE link to the plugin. This is AutoEdit Content Creator Mode, the best AI Premiere Pro plugin that uses Claude AI to automatically edit your videos, cut bad takes, remove silences, and build rough cuts in seconds.” A YouTube Studio screen shows 140.8K views.

Third time this plugin has crossed the audit (#061 in August, #214 four days ago) and the read is stable: the product is real and genuinely Claude-powered, and the ad keeps lying about the same two things. AutoEdit Creator Mode is a shipping Premiere Pro plugin that transcribes your footage, sends the transcript to Claude, and builds a rough cut on the timeline. Now the “free.” The vendor’s pricing page does have a $0-forever tier — it includes four utilities (paste image, silence cutter, sound effects, motion graphics) and, in the page’s own words, no AI editing access. The Claude editing the reel is showing sits on Lite at $9.99/mo (4 videos, 45 minutes of AI editing), Creator at $29.99/mo (15 videos, 2 hours) or Pro at $79.99/mo (40 videos, 5 hours), or as one-time packs at $19.99 for 6 videos and $149 for three months. The “$12k/month” has no owner: no client, no invoice, no channel named — just a screenshot of a 140.8K-view video, which at typical Shorts rates pays out well under a few hundred dollars. And “it adds captions and edits viral shortform” overreaches: the tool’s own product page says it makes a rough cut by removing weak takes and dead space; captions and silence-cutting are things Premiere already does. A useful tool with an honest $10 on-ramp, sold with a fake “free” and an invented income.

Real Claude-powered Premiere plugin with a $0 tier that excludes the AI editing shown; the “free” and the “$12k/month” are the ad, not the product. See #061 and #214.

Confidence

High

Posted by

Eleven Percent LLC

See the original claim →

The claim“There are people making $20–40k/month in niches so boring you’d fall asleep.” Crypto = “45,000 channels fighting over $4 CPMs”; Medicare supplement plans “$44 RPM, highest on the platform, maybe 12 channels”; septic maintenance “$28 per thousand views, ~6 real channels in the entire US”; well water testing “$26 RPM, ~4 channels… 43 million Americans drink from private wells”; commercial insurance “$38 RPM, basically empty”; farm equipment repair “$24 RPM, ~8 channels… every viewer owns $200,000+ in equipment”; HOA law “$32 RPM, effectively zero channels… 74 million Americans live in HOAs.” “10,000 Americans turn 65 every single day.” Funnel: “DM me.” 1K likes, 243 comments, 122 shares.

The thesis — finance, insurance and legal topics pay the highest ad rates and have less entertainment competition — is true, and one stat is exactly right (the EPA does say over 43 million Americans drink from private wells). Everything that makes the reel a money claim is made up. The channel counts collapse on one search: the “effectively zero” HOA niche has a 38K-subscriber story channel launched in mid-2025 with 6.3 million views, plus an attorney-run HOA podcast channel and several drama-story farms; “~4” well-water channels includes one with 47K subscribers after 14 years and a filtration directory listing 50; “~6 septic channels” is contradicted by an article complaining there are too many septic YouTubers; crypto’s “45,000 channels” is off by about 10x. The RPMs are inflated and the units are mixed — crypto gets a “$4 CPM” while everything else gets an “RPM,” and $44 RPM sits above every published 2026 niche range (finance and insurance top out around $32–35 RPM as channel averages). The demographics are stale: 78.1 million Americans live in HOAs per the 2025 industry review, and about 11,400 turn 65 each day, not 10,000. Then the arithmetic: $20–40K a month at even $44 RPM is 455K–909K monetized views every month, and at realistic $20–25 RPMs it’s 0.8–2 million. A niche that only supports four to twelve channels is small precisely because its audience is small — the “empty niche” and the “big check” cannot both be true. The channels that actually earn in Medicare and insurance are licensed agencies using YouTube for lead-gen; AdSense is the rounding error. The credit line points to a co-owner of an AI YouTube-scriptwriting tool, the listicle is a recycled genre, and the creator’s only answer to “how?” is “DM me.”

True premise (high-RPM boring niches exist), fabricated specifics (channel counts, $44 RPM, stale demographics), and arithmetic that can’t produce $20–40K/month — with “DM me” as the product.

Confidence

High

Posted by

Led Strategy

See the original claim →

The claimA split-screen skit: the wanna-be “built an AI agent and let it figure everything out” and “turned it into a multi-agent system so the agents review each other”; the $200K engineer gives the agent narrow, scoped tools, keeps every deterministic step outside the LLM, uses multiple agents only when the task requires it, measures task completion, cost and failure recovery on evaluation datasets, and adds checkpoints with human approval before any high-risk action. Closer: “Join skool.com/baswe-ai … everything you need to land and pass a job interview in the next 90 days.”

Strip the costume and this is Anthropic’s “Building effective agents” post and OpenAI’s practical guide to agents, read aloud. Anthropic: find the simplest solution possible and only add complexity when needed; use predictable workflows before autonomous agents; invest in tool design; have agents pause for human feedback at checkpoints. OpenAI: start with a single agent and evolve to multi-agent only when needed; route irreversible or sensitive actions to human operators; guardrails and evals. The skit’s “$200K engineer” says exactly those things — narrow tools, deterministic steps outside the model, multi-agent only when the task demands it, evals on completion, cost and recovery, human approval before high-risk actions. That is the correct, current consensus, and it’s the most useful 60 seconds in this creator’s funnel. The soft spots are the wrapper: “$200K+” is a costume, not a receipt, and “pass a job interview in the next 90 days” doesn’t match the program’s own 176-day average — nor does knowing these patterns get you through a systems-design loop by itself. The Skool community behind the link is $89/month.

The agent-design advice is correct and matches Anthropic and OpenAI’s own guides; the $200K costume and the 90-day promise are marketing.

Confidence

High

Posted by

Bashiri Smith (Baswe.AI)

See the original claim →

The claim“If I had 6 months to become a $150K–$300K+ AI engineer, this is exactly what I’d learn” — five categories: LLMs & RAG; agents, integration & orchestration (LangGraph, MCP); ops & evaluation; safety & ethics; ML foundations (MLflow, DVC). Closer: “Want to land a $150k+ AI Engineer role before 2027? Reply ‘AI’ to get my full training + the link to join my AI engineering community… everything you need to get hired in 2026.”

The curriculum is fine and the timeline is the industry’s — the numbers are where it stretches. The five buckets are the same sensible syllabus as the creator’s 17-concept reel, and six months is the consensus window for an experienced software engineer to pivot into AI-application work; the community’s own website quotes 176 days average time to an offer, which is six months on the nose. The salary floor is the problem. Self-reported medians for “AI engineer” sit at about $101K (ZipRecruiter) to $142K (Glassdoor); job-posting datasets, which skew senior and San Francisco, put the median near $198K with a 25th percentile around $165K. So $150K is roughly a mid-level engineer at a real tech company — reachable for someone with 3–5 years of software experience who makes the switch, which is exactly who the community says it’s for (“software engineers with 2–5+ years”). $300K+ is senior or staff total compensation, or frontier-lab equity; no dataset shows it as a six-month outcome, and no dataset shows a beginner reaching $150K in six months at all — entry-level base runs $90–135K. The reel shows the plan without the prerequisite: it’s a plan for engineers, sold to a feed of people who mostly aren’t.

Sound curriculum and an honest six-month window for working engineers; the $300K and the beginner-friendly framing are the stretch.

Confidence

High

Posted by

Bashiri Smith (Baswe.AI)

See the original claim →

The claim“If you want to become god-level at AI engineering, learn these 17 concepts” — LLM fundamentals, basic and advanced RAG, agent fundamentals, agentic design patterns, MCP, multi-agent systems, evals, LLM-as-a-judge, agent evaluation, guardrails, observability, LLMOps, CI/CD for LLM apps, data versioning with DVC, ML metrics, and AI governance (NIST AI RMF, EU AI Act). “Comment ‘AI’ and I’ll send you the full roadmap plus a link to my AI engineering community with daily calls with working AI/ML engineers and live guidance from recruiters and hiring managers.” 801 likes, 399 comments.

The list is good. It’s essentially the 2026 AI-engineer syllabus you’d find on roadmap.sh or in the standard AI Engineering textbook: RAG, agents, evals, observability, guardrails, governance — and the video’s own scroll-through shows the roadmap pointing at free public resources (DeepLearning.AI short courses, IBM Technology on YouTube, NVIDIA’s open-source NeMo Guardrails, Langfuse’s MIT-licensed observability, a Coursera responsible-AI course, a vendor’s EU AI Act explainer). Nothing on the roadmap costs money. What the reel doesn’t say is where the comment goes: the community behind it is a Skool group at $89/month (234 members as of today, with the about page noting the price rises at 500), and the “recruiters and hiring managers” on the daily calls are members who work at tech companies, not a placement pipeline. Two honest caveats on the content itself: seventeen bullet points is a table of contents, not “god-level” — there are no projects in it — and it skips prompt/context engineering, fine-tuning, inference cost and prompt-injection security, which any working AI engineer hits in month one. Fair roadmap, free resources, paid funnel.

A legitimate, free-resource roadmap used as bait for an $89/month community — the list holds up; ‘god-level’ and the funnel don’t.

Confidence

High

Posted by

Bashiri Smith (Baswe.AI)

See the original claim →

The claim“I gave Elon’s Grok Bot $1,000 and one rule. Grow it or I delete you. 15 hours later it is holding $3,900. Every trade landed on a public address. No private account and no screenshots.” The bot “reads the news and the filings before the rest of the market… writes the exit before it takes the entry… reads its own losing trades at night and rewrites the rules by morning. Gold, oil, a stock you inherited or a coin that launched an hour ago, the rules are the same. Setup took one evening… Mine costs $300 a month. The whole build is going up soon.” 110K views.

We did what the post dares you to do and pulled the wallet. The address isn’t in this post, but the poster published it two days earlier, and the Solana chain shows a real run: funded with 10.29 SOL (about $1,017) on Aug 25 at 21:57 UTC, then 1,487 transactions in roughly 22 hours — one every 53 seconds, 98 of them failed — across at least eight pump.fun memecoin mints. On Aug 26 at 19:45 UTC the wallet was swept out: 34.18 SOL (about $3,300) moved to another address, leaving a zero balance. So a ~3.3x memecoin-sniping day happened. Everything around it is framing. “It is holding $3,900” was false when posted on Aug 28 — the funds had been gone for two days, and $3,300 isn’t $3,900. “Gold, oil, a stock you inherited” never happened — the chain shows only memecoins, and no single product can trade CME futures, a brokerage account and an hour-old Solana token through one bot. The chain also can’t tell you who drove the trades: Grok Bot, a human, or an off-the-shelf sniper. Grok Bot is real (xAI beta, Aug 11, 2026, bundled with the $300/mo SuperGrok Heavy plan) but has no native trading and requires human approval for financial actions. The reply that asks whether this is “engagement spam to farm the coin” is the key: the poster launched a $CVXV666 token on pump.fun on Aug 23 — “an AI desk that trades memecoins 24/7, 40% of trading revenue to holders” — whose own site says “Paper mode. Live size = 0.” Base rates matter too: in the Alpha Arena live-trading benchmark, Grok 4 lost 45% and four of six frontier models lost money; Bloomberg found only 6 of 32 AI stock-contest runs were profitable.

The wallet is real and did ~3x on memecoins; the “holding $3,900,” the “gold, oil, stocks,” and the missing address are the tell — the product is the poster’s own token.

Confidence

High

Posted by

@antpalkin (cvxv666)

See the original claim →

The claim“Anthropic Engineer Andrej Karpathy just said what no one wants to admit: ‘Agents aren’t magic. They’re distillation at scale.’ ‘Your agent loop is generating garbage at scale and you can’t see it.’ LLMs are wasting 99.99% of their capacity on useless data. In a 20-minute conversation, Karpathy reveals the full reasoning stack. Worth more than any $500 AI course you’ve seen this year.” 91K views, embedded 20-minute video, quote-tweeting a ‘Grok Bot + Kimi K3 300-agent setup’ guide.

The one thing this post gets right, it gets right by accident. Karpathy really did join Anthropic — on May 19, 2026, to the pretraining team — but the video is a cut of the No Priors podcast episode “Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI,” published March 20, 2026, two months before he joined, when he was still running Eureka Labs. A reply under the post names the source. So “Anthropic engineer just said” is a 5-month-old interview with a job title that wasn’t true when he said it. And he didn’t say it: we searched the full 66-minute transcript for “distillation,” “magic,” “garbage,” “waste” and “capacity” — none of the quoted sentences appear. The only “distillation” in the episode is him describing his 200-line microGPT as “a distillation of LLM training,” which is about code, not agents. The “99.99%” number is stitched from elsewhere — his Oct 2025 Dwarkesh interview where he called random internet documents “total garbage” and argued for a small “cognitive core,” and a March 2025 tweet that “99.9% of attention is about to be LLM attention,” which is about documentation, not wasted capacity. What he actually said on No Priors is worth hearing: he hasn’t typed a line of code since December, agent failures are usually “a skill issue” in the harness, and his autoresearch loop found training tweaks overnight that he’d missed. None of that needs a fake quote. The quote-tweet underneath is the point: a crypto-adjacent account’s paid “300-agent setup” guide.

Real interview, real (now) affiliation, fabricated quotes — a cropped March podcast re-captioned to sell an agent-swarm guide.

Confidence

High

Posted by

@0xMortyx (Morty)

See the original claim →

The claim“They built an AI agent in 20 minutes and sold it to Anthropic for $100k — and at Stanford showed how to repeat it from scratch: 02:00 build your first agent with Opus 5 in 20 minutes · 22:34 one agent replaces an Anthropic engineering team worth $200k a year · 49:47 how to make $1.2M a year selling agents to Anthropic and Google. After watching I spent 20 minutes in Claude Code — the agent replaced 10 Anthropic engineers and they paid $100k and asked for more.” 97K views.

This is the previous entry’s template run again 24 hours later with the nouns swapped: “18-year-old / Microsoft / $2.4M” became “they / Anthropic / $100k,” the three timestamps moved, and the quote-tweeted article underneath — the account’s own “Grok Bot + Kimi K3, $100k/month” piece — stayed identical. Anthropic does not buy agents from strangers; its 2026 deals are companies (Vercept in February, Fractional AI and Stainless in May, and reported talks to acquire Decart for about $6B in August). The “$100k” number most plausibly leaks from coverage of Anthropic’s own experiment letting Claude run a small shop. The “one agent replaces an engineering team worth $200k a year” line recycles the exact figure from the day before, and the “$1.2M a year” figure is the account’s house number — a reply under the post: “funny how the math always lands on exactly 1.2m no matter the video.” The video itself is a lecture, not a deal announcement. Replies asking “what’s the actual deliverable they paid for?” never get an answer, because there isn’t one.

No $100k Anthropic deal, no 10 replaced engineers — the same template as the day before with the buyer’s name swapped.

Confidence

High

Posted by

@noisyb0y1 (Noisy)

See the original claim →

The claim“She is 18, built a game with 4 agents and sold it to Microsoft for $2.4M — and at Stanford shared how to repeat it from scratch: 00:09 how Opus 5 in Claude Code built a game making $100k a month · 23:34 4 agents replaced a developer team worth $200k a year · 59:47 how to build your first game from scratch in 60 minutes. After watching I gave Claude Code the game idea I’d been holding for 2 years — in 60 minutes it already existed in the App Store and was making money.” 199.7K views.

There is no record anywhere of Microsoft buying a game from an 18-year-old for $2.4M in 2026 — not in Microsoft’s announcements, not in the trade press, and the top reply under the post says exactly that. The video’s own poster frame gives it away: a Stanford lecture hall, an adult lecturer at a lectern, the university logo in the corner. It is Stanford’s CS329A “Self-Improving AI Agents” course overview — uploaded by the official Stanford Online channel on Aug 3, 2026 (316K views), recorded in September 2025, taught by two Stanford professors (one formerly at Anthropic and Google DeepMind). The lecture covers scaling laws, RLHF and agent loops, cites Claude Code once as an example, and contains no games, no Microsoft, no sale, and no dollar figures. Stanford’s paid “Agentic AI” program built on that curriculum opens enrollment Sept 2 — which is why the footage is being farmed right now. The timestamps and chapter titles are a template — the same account posted “sold to Anthropic for $100k” the next day with different numbers and the same quote-tweeted self-promo article underneath. The closer is impossible on its face: you cannot have an app “in the App Store and making money” 60 minutes after writing it. Apple’s program is $99/yr with identity verification, review averages under 24 hours but often runs days in 2026, and payouts are monthly. The nearest true story this year is a teenager whose satirical Steam game briefly grossed about $1.3M before roughly 99% of it was refunded — no Microsoft, no agents.

No deal, no 18-year-old, no 60-minute App Store launch — a Stanford lecture re-captioned with a made-up sale to farm engagement.

Confidence

High

Posted by

@noisyb0y1 (Noisy)

See the original claim →

The claimA 7-slide carousel: ERPNext (38k stars) gives you “invoicing, stock and payroll in one free install,” Plane (57k) replaces Jira/Linear/Monday/ClickUp “with no per-seat pricing,” Trigger.dev (16k, Apache 2.0) “runs your recurring jobs so the invoices go out at 3am without you opening a laptop,” Lago (10k) meters usage into “pricing, credits, invoices and payments,” and Relaticle (1.5k) is “the CRM built for right now” with a 32-tool MCP server. Closer: “Stand up ERPNext this weekend and see how much of your stack it swallows.”

This poster’s 15th appearance on the audit (14 prior entries), and the pattern holds: the repos are real and the numbers are honest, the verbs are not. We pulled every repo from the GitHub API on Aug 29 — ERPNext 38,635 stars, Plane 58,497, Trigger.dev 16,156, Lago 10,442, Relaticle 1,566 — so the slides are accurate to the rounding. The oversell is in what they do. Payroll is not in ERPNext; it was split out into the separate frappe/hrms app at version 14, a second install. Plane’s cloud is entirely per-seat ($6 and $13 per seat per month) — “no per-seat pricing” only holds if you self-host the AGPL community build. Lago’s payment-provider integrations and dunning sit on the Premium tier. And Trigger.dev’s own self-hosting guide says the words “unlikely to result in a production-ready deployment” and asks for two machines plus Postgres, Redis, ClickHouse and object storage. Nothing here “runs a side hustle automatically” — you would be the unpaid sysadmin for roughly four databases and five Docker stacks, and the CRM you’d bet your pipeline on has one real contributor.

Five real repos with honest star counts — but no payroll in ERPNext, per-seat pricing at Plane, Premium-gated payments at Lago, and a self-host burden the tools’ own docs warn you about.

Confidence

High

Posted by

@ty.prompts.ai (Ty AI)

See the original claim →

The claimA carousel of 7 prompts that “actually make you use what you learn” — spaced-repetition cards, a 48-hour real-world application plan, and more — closing with: “same prompt, every model, side by side. Search ‘Chappie’ on the App Store.”

This one is refreshingly honest: the study techniques are genuinely evidence-backed. Spaced (distributed) practice and retrieval practice — exactly what the “spaced-repetition cards” and “quiz yourself” prompts do — are the two highest-rated learning techniques in Dunlosky et al. (2013), the big review in Psychological Science in the Public Interest, and the “apply it within 48 hours” prompt is reasonable transfer practice. The app is real too: ‘Chappie’ is on the App Store and does what the closer says — run one prompt across ChatGPT, Claude, Gemini, Grok and DeepSeek side by side. The only soft spot is the ad quietly leaving out that there’s no free tier ($99.99/year behind a 7-day trial) — and that you don’t need the app at all: the prompts work in any free chatbot.

Real, evidence-backed study prompts that are free to use anywhere — the app they’re selling is a $99.99/year convenience, not a requirement.

Confidence

High

Posted by

@chappie.us (chappie)

See the original claim →

The claimA viral post quoting the “Head of Claude Code” saying “I’m not prompting my agents anymore, I’m building graphs and loops so they can build the agents for me,” promoting a 10-minute talk and a step-by-step “graph engineering” guide, with the tagline “most people use AI at 5–10% of what it can do.”

There’s a real person and a real idea underneath the polish. Boris Cherny genuinely is the creator and head of Claude Code, and he really did say (June 2026, covered by The New Stack, Fortune and others) “I don’t prompt Claude anymore. I have loops running that prompt Claude … my job is to write loops.” That shift — from hand-typing prompts to running standing agent loops — is real and worth understanding. But the post embellishes: Cherny said LOOPS; “graphs” and “so they can build the agents for me” are spliced in from a separately-trending “graph engineering” meme he’s not on record saying. “Graph engineering” itself is a mid-2026 community label slapped onto multi-agent patterns Anthropic already published — real technique, manufactured branding. And “5–10% of what AI can do” is the same floating, unmeasurable hook this account reuses to funnel readers to its guides.

Cherny really runs loops instead of hand-prompting — that’s real. The doctored ‘graphs’ quote and the buzzword are the growth-account gloss.

Confidence

Medium

Posted by

@AnatoliKopadze (Anatoli Kopadze)

See the original claim →

The claimA post pairing Geoffrey Hinton’s “if you sleep well tonight you may not have understood this lecture” with the claims that “most people use maybe 10% of Claude” and “I turned his talk into 17 Claude features 99% of users never find.”

The ingredients are real; the assembly is bait. Hinton’s lecture is genuine — a Royal Institution Discourse, roughly 47 minutes, and it really does open with that “sleep well tonight” line (though it’s from around October 2025, not ‘recent’). Hinton is a real, Nobel-winning pioneer who did leave Google to warn about AI. But “already ahead of us on most cognitive tasks” overstates what he actually argues (AI is heading past us, ahead on some things like raw knowledge breadth), and “built the neural networks behind every AI alive” is hero-myth compression — he’s one of several founders. The tell: this exact post text is a recycled viral template that this account and several other growth accounts have posted word-for-word, and Hinton’s AI-safety lecture has nothing to do with the “17 Claude features” it’s glued to — which are all standard, documented Claude features. “Use only 10% of Claude” is unfalsifiable.

Real lecture, real features — bolted together with an overstated quote and a recycled template to feed a newsletter.

Confidence

Medium

Posted by

@AnatoliKopadze (Anatoli Kopadze)

See the original claim →

The claimAn ad implying you can charge clients $16,300 a month to edit videos using Adobe Premiere Pro plus an AI assistant (Claude), via the poster’s plugin.

The product is real and genuinely Claude-powered — which is rarer than you’d think. Eleven Percent’s AutoEdit (Creator Mode) is a real, shipping Adobe Premiere Pro plugin that transcribes your footage, sends the transcript to Claude, and builds a rough cut on the timeline for you. Pricing runs from roughly a $10 preorder up to $180–$450 for the full plugin collection. But the vendor’s own site says the output is a rough cut, not a finished edit — and the “$16.3K/mo” number appears nowhere on it: no client names, no receipts. Real editing retainers mostly run $750–$5,000/mo; $16,300 is a top-fraction-of-1% outlier, and the oddly precise “.3” is classic credibility theater. “Claude edits your video” is the false leap — an LLM assembles a rough cut from text; a human still makes the actual edit.

Real Claude-powered Premiere plugin that speeds up rough cuts; the $16.3K/mo client figure is an unverified advertising hook.

Confidence

Medium

Posted by

Eleven Percent LLC

See the original claim →

The claimA reel showing a personal ‘Jarvis’ AI assistant phoning the driver with updates — implying an ambient, finished AI concierge you could just have.

This one is real capability wearing a movie poster. An AI that phone-calls you with updates is a completely standard build today — an outbound voice agent (Vapi, Retell, Bland, or Twilio wired to an LLM and a text-to-speech voice) fired by a webhook or a schedule. So it’s plausible-real, not special effects. But as a reel it’s a scripted demo, and it’s a software engineer’s custom rig — not a product you can download. Luke Cutting runs it as a funnel: a free ‘JARVIS crash course’ that feeds a $97/month community, with revenue claims (“it 3x’d my revenue”) that are unverifiable self-reports.

The outbound-voice-agent tech is real and we build it — but this is a custom rig selling a $97/mo course, not a downloadable Jarvis.

Confidence

Medium

Posted by

Luke Builds AI (Luke Cutting)

See the original claim →

The claimA reel: “If I went broke, give me Claude Max, FanProStudio.ai, and 2 AI girls — I’ll be up $50K in 4 weeks,” pitching AI-generated influencer/‘girlfriend’ personas as a fast income play.

The tools are real; the math and the plan are not. FanPro Studio is a real AI-influencer platform (launched mid-2026, identity-locked AI personas), and Claude Max is a real $100–$200/month subscription. But “$50K in 4 weeks” has zero documented cases: the most famous AI persona on earth (Aitana López, built by a professional agency over years) averages around €3,000/month, peaking near €10–11K — this claim is roughly 15× that, from a standing start, with no methodology, no named account and no receipts. It’s fabricated income bait. It also breaks the rules it name-drops: Claude generates no images or video (it’s irrelevant here beyond captions), and Anthropic’s usage policy prohibits sexual content and presenting AI output as a real human — the two things this play actually runs on. OnlyFans bans fully-synthetic personas outright; only Fanvue allows them, and only with a verified human behind the account and mandatory AI disclosure.

Real tools, fabricated money, and a plan that breaks Anthropic’s usage policy — straight BS.

Confidence

High

Posted by

Simone (AI Influencer Monetization)

See the original claim →

The claimA red/green traffic-light rating of business mentors for blue-collar owners: Grant Cardone = SKIP, a leadership coach and a local-SEO-for-home-services creator = WATCH. Caption: “Do you agree??”

This is opinion content, not a spec claim, so it can’t be “false” the way a product boast can — but the underlying calls are mostly defensible and one guardrail matters. The Cardone = skip verdict has real weight behind it: Cardone Capital is facing a certified securities class action (Pino v. Cardone Capital), revived by the 9th Circuit in 2025, alleging he promoted “north of 15%” returns on social media that regulators had already questioned. The accuracy guardrail: there is NO SEC fine or fraud finding — it’s a private investor suit that is still unresolved, so the commenters calling him a flat “scammer” are voicing opinion, not adjudicated fact. The red/green traffic-light format is the real weakness: rating a whole human WATCH or SKIP flattens the nuance that actually helps an owner decide who to learn from.

Defensible opinions, real litigation behind the Cardone call — but it’s a take, not a verdict, and “scammer” overstates an unresolved case.

Confidence

Medium

Posted by

@faiezranaofficial (faiez.rana)

See the original claim →

The claim“Holy cow dude Elon really cook with this. Are you gonna use this?” over Grok footage. #grok #elon #grokbot

The hashtags point at something real — Grok Bot launched in beta August 11, 2026 and Grok 4.6 landed August 12 — but the reel itself makes no verifiable claim at all. No feature is named, no demo is shown, no capability is described. “Holy cow Elon really cooked” is pure keyword-bait riding Musk’s name to farm the algorithm. There’s nothing here to be right or wrong about, which is exactly the problem: contentless hype trains you to react to a name instead of evaluating a tool.

Real product underneath, but the clip is a name-drop with no claim — hype with the information deleted.

Confidence

High

Posted by

@lord.luke5

See the original claim →

The claimA breakdown of a multi-monitor autonomous-agent setup (“I’m building Ultron”), explicitly crediting “Original video by Sagar Tamang.”

This is the rare repost done right. The underlying project is real and documented — Sagar Tamang’s U.L.T.R.O.N.: a holographic Three.js orb with gesture and voice control plus an Android device-control agent layer (the orb is open-source, the control layer proprietary). Crucially, Luke credits the original creator by name, which defuses the usual stolen-content problem entirely. The “Ultron vs Jarvis/Friday” framing is branding flourish, but there’s genuine technical substance under it, and nothing false is asserted. When a breakdown names its source and describes a real build accurately, it holds up.

Properly attributed, technically real, no false claim — it holds up.

Confidence

High

Posted by

Luke Builds AI

See the original claim →

The claimA drifting glowing-particle sphere built in Three.js, with real HTML/JS source on screen. Caption: “Tiny points of light just drifting around. #threejs”

This is a legitimate Three.js aesthetic post — the visible source (a createParticles function driving a BufferGeometry point cloud) produces exactly the drifting-points effect shown. There is no product, no income claim, and nothing oversold; it’s a code demo. The only soft flag is authorship — commenters note there’s no link in the bio and ask whether the poster understands the code — but nothing indicates a scam. It holds up as what it is: low-substance, honest eye-candy.

An honest code demo with no false claim — it holds up.

Confidence

High

Posted by

Coding Withme

See the original claim →

The claim“POV: you finally mastered 3D websites. Comment WORK and I’ll send the exact workflow: Concept → Design → Development → Transfer (test across devices, optimize load times).”

Concept → Design → Development → Transfer is a real and correct high-level pipeline — but at that altitude it describes literally any website build, not anything 3D-specific. There’s no mention of Three.js or WebGL, GLTF asset budgets, draw-call limits, level-of-detail, or the scroll libraries where the actual difficulty of 3D sites lives. The visible content is generic filler; the real value (if any) sits behind the “comment WORK” DM-gate, which is standard reach farming. The framework is defensible but thin, used as bait rather than taught.

A correct pipeline at the wrong altitude — true, generic, and gated as bait.

Confidence

Medium

Posted by

Gilt Foundry

See the original claim →

The claim“No Claude was used. The scroll video was generated with ByteDance’s new Seedance 2.5, available only on Higgsfield AI. Comment VIDEO for the prompt.”

Seedance 2.5 is a real ByteDance model (launched July 31, 2026 — 4K, synced audio, up to 30 seconds). But “available only on Higgsfield AI” is false: ByteDance shipped it first on its own Jimeng AI and Doubao apps with Volcano Engine Ark API access, and multiple third-party platforms carry it — Higgsfield’s own blog even runs a “best platforms to access Seedance 2.5” roundup. The single-platform push plus a DM-gate fits the affiliate pattern flagged in prior audits (Higgsfield runs an ambassador/affiliate program). “No Claude was used” is an irrelevant flex — Claude isn’t a video tool. Real model, misleading funnel.

Real ByteDance model, false “only on Higgsfield” exclusivity — an affiliate hook.

Confidence

High

Posted by

Sean - UX Designer

See the original claim →

The claimSix AI video generators that are “actually free” and run on a “consumer GPU,” shown via an open-source video-model repo.

The models are genuinely free and open-source — Wan 2.1/2.2, LTX-Video, Mochi 1, HunyuanVideo, CogVideoX, AnimateDiff easily clear six — so “actually free” is true on licensing. But the commenters have the real point: “consumer GPU” is doing heavy lifting. Only LTX-Video, CogVideoX, and AnimateDiff run comfortably on an 8–16GB card; Wan wants ~24GB and HunyuanVideo (13B params) realistically needs 40–80GB or rented cloud. The software costs nothing; the hardware to run the good ones isn’t “consumer.” One commenter nailed it: “‘consumer GPU,’ then proceeds to mention 11B [params].”

Genuinely free models, misleading “consumer GPU” framing — the cinematic ones need a workstation or the cloud.

Confidence

Medium

Posted by

Eric Socal

See the original claim →

The claimBoris Cherny, creator of Claude Code at Anthropic, gave a 30-minute workshop of 7 tips; Anthropic’s technical onboarding went from 2–3 weeks to about 2–3 days using these practices.

Almost entirely accurate. Boris Cherny is the verified creator and head of Claude Code at Anthropic, and the seven tips match Anthropic’s official “Mastering Claude Code in 30 Minutes” workshop nearly verbatim: start with codebase Q&A, practice prompting, teach it your team’s CLI tools and MCP servers, match workflow to task, give it more context, tune that context, and set up CLAUDE.md plus MCP, permissions, and slash commands checked into git. The “2–3 weeks to 2–3 days” onboarding line is a direct quote from that workshop. The tips are sound, standard best practices — nothing to debunk.

Real person, real workshop, accurate quote — it holds up.

Confidence

High

Posted by

Evolving AI

See the original claim →

The claim“Had no coding experience before making this, I’m a finance major. Roblox development has never been easier — check out my game to see what YOU could do in under a month.”

That an individual built a simple game with no prior experience is plausible and fine — Roblox Studio genuinely lowers the barrier. But the implied “anyone can make money” is survivorship bias. In 2025 Roblox paid out roughly $1.5B to creators, yet the distribution is brutally skewed: an estimated 60% of all earnings go to the top ~500 experiences, and only a small fraction of the tens of thousands of developers ever qualify for or execute a payout. The overwhelming majority earn essentially nothing. “No coding, easy money in a month” hides that a tiny elite captures nearly all the money.

Real that a beginner can build; false that it’s an income path — a tiny elite takes nearly all the money.

Confidence

High

Posted by

@bennettbenito

See the original claim →

The claimThree AI tools “actually boost your video engagement,” credentialed by a “$6k/month SaaS founder”: RetentionMaxAI (second-by-second drop-off analysis), VidIQ (channel audit + ideas), Opus Clip (long-form to shorts).

VidIQ and Opus Clip are real, established, widely used tools with fair claims — Opus Clip’s “10 clips from one video” is literally its function. RetentionMaxAI, however, doesn’t surface as any established product; combined with the #retentionmaxai tag, it reads as the poster’s own tool slipped into a list of legit names to borrow their credibility. That’s undisclosed self-promotion, and the unverifiable “$6k/month SaaS founder” line is classic authority-padding. Trust the two known tools on their merits; treat the boast and the unnamed-until-hashtag product as advertising.

Two real tools carrying an undisclosed self-plug and an unverifiable founder boast.

Confidence

Medium

Posted by

@content.wizard

See the original claim →

The claim“Claude doesn’t know social media, but connect Sandcastles AI (by Kallaway) and Blotato (by Sabrina Ramonov) and it becomes a full social-media powerhouse — formats, scripts, posting across all platforms.”

Both products are real and legitimate. Sandcastles AI (from creator Kane Kallaway) is a viral-script and outlier-research tool; Blotato (Sabrina Ramonov) is a real social scheduling/publishing service around $29/month that genuinely connects to Claude via a hosted MCP server. The overselling is “turn Claude into a powerhouse that manages your social.” These are two narrow tools — idea/scripting plus scheduling — and the strategy, editing, brand judgment, and community engagement still aren’t automated. “Claude sucks for social media” is a hook, not a finding.

Two real, legitimate tools oversold as a hands-off creative team.

Confidence

Medium

Posted by

Justyn The AI Guy

See the original claim →

The claim“AI agents now write, design, and schedule an ENTIRE MONTH of content while you sleep. Here’s the 7-prompt system that turns Claude into your full-time creative team.”

Same genre this audit has graded before, same verdict. The prompt system — content architect, voice clone, and so on — is real and can genuinely draft a month of posts fast. But “while you sleep” oversells the autonomy. AI drafts still need a human to edit for voice, fact-check, approve, design the final assets, and answer comments and DMs — which is most of what a social media manager actually does. It’s a drafting accelerator, not an autopilot, and “stop hiring” is the hook that makes it sound like a replacement.

Real drafting prompts, oversold as a hands-off content team.

Confidence

Medium

Posted by

@boscokat

See the original claim →

The claim“These 5 prompts quietly replace $500/hour consultants — most people don’t realize ChatGPT can do this,” starting with “Roast my resume” (a brutal FAANG-recruiter critique).

The prompts are genuinely useful and free — a “brutal FAANG recruiter” critique of your resume has real value, and so do negotiation-prep and similar templates. But a prompt doesn’t replace a $500/hour consultant. A consultant brings accountability, context tailored to your specific situation, industry relationships, and judgment about your particular case that a one-shot generic prompt can’t. The “quietly / most people don’t realize” framing is standard engagement bait wrapped around real utility.

Genuinely useful prompts, oversold as a consultant replacement.

Confidence

Medium

Posted by

@betterpromptsai

See the original claim →

The claim“8 Claude AI side hustles that replace a full-time job,” tagged #HiringNow #jobs #nowhiring.

Tagging a list of solo income ideas with #HiringNow, #jobs, and #nowhiring is deliberate misframing — nobody is hiring you; you’re being sold unproven self-employment dressed as job openings to farm the jobs-search algorithm. “Replace a full-time job” implies stable, replicable income, but a slideshow of generic Claude prompts is a list of possibilities, not a paycheck — and most people who try these earn close to nothing. The hashtags are the tell: they exist to catch people searching for actual work.

Self-employment ideas mislabeled with hiring hashtags to farm job-seekers — the framing is the scam.

Confidence

High

Posted by

@timothy.dew3

See the original claim →

The claim“Automaton is an AI that survives by making money and clones itself if profitable — a viral new invention showing how AI could run entire businesses autonomously.” #cryptoai

Automaton is a real open-source project, not vaporware: built by Sigil Wen / Conway Research, it gives an AI agent its own crypto wallet (USDC via the x402 payment protocol) so it must “earn its existence” by building and selling products and paying for its own compute — die if it can’t cover costs, replicate if profitable. So the literal description is accurate. Why it’s not “real”: it’s a provocative research demo and crypto-flavored art project, not a proven autonomous business engine. There’s no evidence it actually sustains profitable self-replicating businesses, and even Vitalik Buterin publicly criticized it as off course. The “#cryptoai” viral packaging is the tell; “AI could run entire businesses autonomously” is aspiration, not a shipped result.

A real, viral crypto research demo — not the autonomous business engine the caption implies.

Confidence

Medium

Posted by

@harrison.about.money

See the original claim →

The claim“FINALLY, AI can SELL instead of you — Vladimir built the first AI sales tool that does the whole job.” Explee’s AutoGTM, a 24/7 outreach agent.

AutoGTM is a real, launched product: paste your domain and about seven AI agents research the market, build ICPs, find and verify contacts, write personalized cold emails, and send on a schedule at roughly $0.03/email with free starter credits. The founders are real. But the marketing is the same claim this audit keeps grading BS: “FIRST AI sales tool that does the whole job” is false — Clay, Artisan, 11x, Jason AI and dozens more predate it — and “AI can SELL instead of you” is automated top-of-funnel outreach, not selling. It doesn’t discover needs, handle objections, build trust, or close. High-volume cold email also risks deliverability and brand damage. Real tool, false superlatives.

Real outbound-automation tool, false “first / replaces your salesperson” superlatives.

Confidence

High

Posted by

Explee AI

See the original claim →

The claim“NVIDIA’s Nemotron 3 Ultra is a 550B-parameter MoE model built for coding — try it FREE on OpenRouter/OpenCode. I gave it one prompt and it built a full 3D landing page.”

The model is real and the numbers are exact: NVIDIA-Nemotron-3-Ultra-550B-A55B — 550B total, 55B active MoE — and it’s genuinely free on OpenRouter and usable via open coding agents. So the factual core is solid. The hype is the framing: “just came for Claude Code” is creator language, not a benchmark result — a free open-weights model doesn’t equal beating Claude on agentic coding, and no head-to-head is shown. “One prompt built a full 3D landing page” is a cherry-picked demo; any capable model spits out a Three.js template from one prompt. Real model, worth trying free — but “killed Claude Code” is clickbait.

Real, free NVIDIA model wrapped in an unearned “killed Claude Code” headline.

Confidence

High

Posted by

@ai.amirtech

See the original claim →

The claim“The Autonomous Intern 2 is a tiny pyramid-shaped computer built to run AI agents — connect it to Slack/Telegram/Discord, give tasks by talking, keep memory/files/API keys local. Preloaded with OpenClaw or Hermes.”

This is a genuine shipping product from a company called Autonomous: a ~4.7-inch glowing pyramid desktop computer built to run AI agents locally, connect to Slack/Telegram/Discord, take voice or text tasks, and keep memory, files, and API keys on-device. The specs check out — an Allwinner A733 8-core SoC, a 3-TOPS NPU, open-source “Autonomous OS,” preloaded with OpenClaw or Hermes — and pricing is real at $249 for the OpenClaw/Hermes editions and $299 for the Developer Edition. Matt Wolfe’s description is accurate. The one honest caveat: 3 TOPS is modest, so it can’t run large frontier models on-device; it’s an always-on orchestration box that mostly calls cloud APIs (hence “store API keys locally”).

A real, shipping $249–$299 product — accurately described, with a modest on-device ceiling.

Confidence

High

Posted by

@mattrwolfe (Matt Wolfe)

See the original claim →

The claimSearchPhone runs phone lookups across Google/DuckDuckGo/Reddit/GitHub at once, validates numbers + carrier via Numverify, checks the Hudson Rock infostealer database, and auto-generates JSON + PDF reports — only 3 free API keys required.

The repo genuinely exists and does roughly what’s claimed: a Python CLI OSINT tool that validates numbers and pulls carrier via Numverify, runs parallel searches across Google/DuckDuckGo/Reddit/GitHub, checks the Hudson Rock Cavalier infostealer database, and outputs JSON + PDF reports. The “3 free API keys” is roughly accurate — the Hudson Rock check needs no key, so the paid surface is small. It’s a functional aggregator, not magic; results are only as good as what’s already public, and dorking GitHub for a phone number is hit-or-miss. The important flag for Schreier Group: this is surveillance-adjacent tooling — running lookups on real people without consent raises stalking, privacy-law, and breach-data concerns, and the infostealer angle is literally breach data. Real and accurate, but not a tool to attach to a client engagement.

A real, functional OSINT tool — accurate claim, but surveillance-adjacent and off-limits for client work.

Confidence

Medium

Posted by

@whitewhoadie

See the original claim →

The claim“Comment if you want the download in the Discord!” — a Claude Code + Hermes agent setup gated behind joining a Discord. #claude #claudecode #hermes

Hermes is a real framework — NousResearch’s hermes-agent (~96k stars) — and a community port (hermes-CCC) genuinely brings it into Claude Code’s skill system with dozens of skills and no OAuth. So the underlying “Claude Code + Hermes” tool exists and is free and open-source on GitHub right now. The problem is the delivery: gating a public, permissively-licensed repo behind “comment → join my Discord → get the download” is a classic audience-funnel play, not a necessary distribution step — anyone can clone it directly. (Note: a prior audit’s “~231k stars” figure for Hermes is stale; it’s ~96k now.)

Real free tool, unnecessary Discord gate — the funnel is the point, not the download.

Confidence

Medium

Posted by

@l_ogannn

See the original claim →

The claim“Ten weeks ago the agent harness barely existed. DeepSeek Harness broke the internet: over 170,000 stars and 19,000 forks in its first week, everything is a plugin.”

The repo is real — deepseek-ai/deepseek-harness, MIT-licensed, launched August 13, 2026, “everything is a plugin” (an accurate quote) — and it genuinely broke GitHub’s fastest-star record (about 20k stars in an hour). But the specific “over 170,000 stars in its first week” overshoots: sources put it around 135–140k by the end of the first week, and the 19,000-forks figure isn’t corroborated. This is Ty Blackard’s recurring pattern — a real, impressive underlying event with the numbers rounded up past what the data supports. The honest version (“~140k stars in its first week, a record-breaking launch”) is still remarkable.

Real record-breaking repo, star count rounded up past the data — a Ty Blackard signature.

Confidence

Medium

Posted by

@ty.prompts.ai (Ty Blackard)

See the original claim →

The claimFive open-source frameworks = five sellable AI-agent businesses: AI receptionist on LiveKit ($1–3k/mo), support agent on Dify, lead machine on Firecrawl, ops automator on n8n ($1–3k/mo retainers), agent SaaS on CrewAI.

The frameworks are all real and roughly the stated size — Dify (~153k stars), n8n (200k+), Firecrawl (~137k), LiveKit’s Agents framework, CrewAI (though it’s ~44–48k, not the claimed 55k). The business models are legitimate too. The overselling is the income: $1–3k/mo retainers for AI receptionists or ops automation are achievable but are the TOP of the range for operators who can actually sell and retain SMB clients — not a default outcome. The AI-generated imagery is a tell that the “results” are illustrative, not real case studies. Real tools, real model, but the money is a ceiling, not a floor.

Real frameworks and business models, with top-of-range income sold as the default.

Confidence

Medium

Posted by

@ai.blueprint

See the original claim →

The claimEight trending repos that replace paid subscriptions for practically free: cloudflare/computer, addyosmani/agent-skills (86.7k), firecrawl/pdf-inspector, vitali87/code-graph-rag, esengine/DeepSeek-Reasonix (34.2k), virgilioj94/book-to-skill, google/skills, zhaoxuya520/reverse-skill.

Unusually clean for this genre. The three repos most likely to be fabricated all checked out: addyosmani/agent-skills is real at ~89k stars (the 86.7k claim is accurate), esengine/DeepSeek-Reasonix is a real DeepSeek-native terminal coding agent at ~35k (34.2k is accurate), and google/skills is Google’s official Agent Skills repo from Cloud Next 2026. Also confirmed real: cloudflare/computer (“give your agent a computer”) and firecrawl/pdf-inspector (a Rust PDF library). No fabricated repos, and the star counts are close to actual. The only honest caveat is the usual “free” framing — free software still needs setup and, for some, cloud costs — but the list itself is legit.

A genuinely accurate repo listicle — real repos, honest star counts.

Confidence

Medium

Posted by

@melibuilds.io (Zach Melinger)

See the original claim →

The claimTen repos with big star counts: mattpocock/skills (203k), anomalyco/opencode (193k, “Claude Code rival”), yt-dlp, garrytan/gstack (126k), unclecode/crawl4ai (76.6k), jamiepine/voicebox (49k), heygen-com/hyperframes, impeccable, microsoft/SkillOpt, trailofbits/skills (“40+ security skills”).

The repos are overwhelmingly real; the inflation is modest, not fabricated. crawl4ai is real but sits around ~60k, not 76.6k. opencode is real and genuinely a Claude Code rival, but ~161–180k, not 193k — and the “anomalyco/opencode” owner is defensible (the project began as sst/opencode and migrated). voicebox (jamiepine, 7 local TTS engines), mattpocock/skills, yt-dlp, garrytan/gstack (Garry Tan’s 23-tool setup, verbatim), heygen-com/hyperframes, and trailofbits/skills all exist as described. Honest corrections: knock crawl4ai to ~60k and opencode to ~170k, and trailofbits shows roughly a dozen-plus skills, so “40+” may be padded. Good list, a few numbers rounded up.

Real repos with a few star counts padded 15–30% — solid list, inflated numbers.

Confidence

Medium

Posted by

@melibuilds.io (Zach Melinger)

See the original claim →

The claim“This AI method is a game changer — comment ‘Claude’ and get the free guide.” The actual method is never stated on screen.

There’s no verifiable claim to grade — the “AI method” is never stated. “7 months of using Claude and NOW I find this?!” plus “comment ‘Claude’” is a pure comment-gate lead-gen funnel engineered to farm engagement and route people into a side-hustle DM sequence. Claude is name-dropped as bait; nothing about a specific capability is asserted or falsifiable. A real method gets shown in the video; a funnel hides it behind a keyword.

No claim, no method — pure comment-gate lead-gen using Claude as bait.

Confidence

High

Posted by

Andrea Creates AI

See the original claim →

The claim“POV: you got Claude Fable 5 to build you a $50k website. Comment HOW for the exact framework I use to create 3D websites of this quality and get high-ticket clients.”

Claude Fable 5 is a real model — Anthropic’s current flagship — so the name is legit. The outcome claim is inflated: Fable 5 can one-shot a strong, polished landing page, but a “$50k-quality” bespoke 3D agency site rarely comes out production-ready in a single prompt; it needs asset work, performance tuning, and iteration. “$50k” is a value-anchoring number, and the “comment HOW for the framework” gate is the standard lead-capture funnel, not a deliverable. Real tool, oversold result, funnel attached.

Real model, inflated one-shot outcome, standard DM-gate funnel.

Confidence

Medium

Posted by

Brandon the Web Dev

See the original claim →

The claimAn open-source “Claude SEO” skill for Claude Code — 25 sub-skills, 18 parallel agents, v2.2.0, 326 tests, MIT licensed — presented with a red overlay reading “CLAUDE SEO TEAM OFFICIALLY!”

The repo is real and genuinely good: github.com/AgricIDaniel/claude-seo — 25 sub-skills, 18 sub-agents, MIT licensed, ~410 tests across the v2 line, audits technical SEO, schema, E-E-A-T, GEO, and local SEO. It’s built by Agrici Daniel, an independent developer — a single maintainer, explicitly NOT Anthropic. The commenters are right: “Claude didn’t release this.” The red “CLAUDE SEO TEAM OFFICIALLY!” overlay is fabricated — there is no such Anthropic team, and the repo itself makes no official claim. Good free tool, false official framing.

Real, good open-source SEO skill with a fabricated ‘official Claude team’ badge.

Confidence

High

Posted by

Wassim younes AI

See the original claim →

The claim“Anthropic quietly created a new job. Claude for Small Business ships 15 pre-built workflows — install it for local businesses, charge £1,500 setup + £500/mo. Five clients = £30K/year.”

The product is real and official: Anthropic launched Claude for Small Business on May 13, 2026 — connectors plus about 15 ready-to-run agentic workflows (invoice chasing, payroll planning, month-end close, margin analysis, contract review) that plug into QuickBooks, PayPal, HubSpot, DocuSign and more. So the earlier ‘it’s made up’ read is now outdated — this ships. What’s oversold is the grift layer: it’s included at no extra charge in a standard Pro/Max subscription (~$20–$200/mo) and it’s designed to be self-serve and light to set up — which undercuts the “they’ll never do it themselves, charge £1,500 + £500/mo” premise. You CAN charge for setup and training as legitimate advisory work; you can’t honestly resell it as a secret product.

Real official Anthropic product; the ‘quietly created a new job’ resale pitch is arbitrage on an included feature.

Confidence

High

Posted by

@nicksadler.io

See the original claim →

The claimUse Claude to scrape Google Maps for blue-collar businesses with no website, cold-call them saying you already built them a free homepage, have LandingSite AI actually build it, then charge $800 setup plus $50/month — a $10K/month side hustle.

The product half is real. LandingSite is a genuine AI website builder (free build-and-preview tier, $10–$20/month per published site, a $49/month white-label agency plan), and its reseller path is real too: a handover model where the platform bills your client through Stripe and takes 5%, or you bill direct and keep everything. Pre-building a preview site before you ever call is genuinely possible on the free tier — and no-website local businesses really are a large, underserved market. Now the other half. “Claude scrapes Google Maps” describes a Google Terms-of-Service violation; the compliant route is the paid Places API, whose “no website” signal is unreliable anyway. The $10K/month needs roughly 12–13 closes a month at $800 against cold-call benchmarks where 2–3% of dials produce a MEETING, not a sale — that’s thousands of dials a month, a boiler-room job wearing a side-hustle costume. The $50/month “recurring” layer nets $24–$38 after per-site platform costs and churns hard with reluctantly-acquired clients. And the script itself — “I already built you a FREE homepage” followed by an $800 invoice — is deceptive-offer framing the FTC has a name for, pitched at cell phones the TCPA still covers. Follow the money: this strategy video comes from the vendor’s own partner channel, and the vendor collects its subscription whether you ever close a single client.

Real builder, real reseller program, invented math — and the ‘free site’ cold-call script is a deceptive-practice complaint waiting to happen.

Confidence

Medium

Posted by

a website-builder company’s partner account, cross-promoted by a hustle-content creator

See the original claim →

The claimConnect Claude Cowork to Zoom — it grabs your call transcripts and auto-creates 10 social posts plus a blog post from every call. Comment “PROMPT” and I’ll send you the prompt.

The plumbing checks out. Zoom shipped an official “Zoom for Claude” connector in April 2026 — it really does pull your meeting transcripts, AI summaries, and recordings into Claude by natural language — and Claude Cowork’s scheduled tasks (recurring, cloud-run since July) can genuinely be set to draft a batch of posts and a blog from your latest calls. That workflow exists and it’s useful. What the reel sells past you: transcripts only exist on a PAID Zoom plan with cloud recording and Smart Recording or Meeting Summary switched on (an admin toggle on managed accounts); Cowork’s scheduled tasks need a paid Claude plan; you can only mine calls you host or were granted access to; and — the big one — Claude drafts files, it does not publish anything. There is no native Facebook, Instagram, or LinkedIn posting in Claude; a human or a separate scheduler still ships every post. “Auto-creates from every call” quietly conflates drafting with publishing, and nobody in the reel mentions that piping client calls into a content machine without the other participants’ consent is a confidentiality problem, not a growth hack. Ten posts per call is a volume promise, not a quality one — and the comment-“PROMPT” gate is a ManyChat-style list harvester, which is the actual business here.

Real plumbing, oversold autopilot — the reel sells a list, not a workflow.

Confidence

High

Posted by

a marketing-agency brand account running a comment-“PROMPT” DM funnel

See the original claim →

The claimHow to get rid of Claude's new invisible watermarks so the things you make aren't AI-detectable — “if you want to use Claude properly and not get watermarked, use this Claude skill to get rid of it.”

The premise is real news, and that's what makes this one dangerous: as of early August 2026, Anthropic does embed an imperceptible watermark in Claude-generated text at the model level — across claude.ai, the API, and Claude Code — plus digitally signed C2PA provenance metadata on supported files. Anthropic's own help center confirms it, with no opt-out; Fortune, Forbes and Gizmodo describe a SynthID-Text-style statistical bias in word choices under a key Anthropic holds, driven largely by EU AI Act Article 50. Now the twist that kills the product: a Claude skill runs INSIDE Claude. The watermark is applied at generation time, inside the model — so any “de-watermarking” rewrite the skill produces is itself freshly watermarked on the way out. You cannot use Claude to wash Claude's watermark out of Claude's output. What actually degrades the signal, per Anthropic's own limitations note, is heavy HUMAN editing, paraphrasing, translation, or mixing with other writing — none of which a skill provides. At best it could strip C2PA metadata from files, which any exif tool does and which tampering the C2PA signature is built to reveal. And Anthropic notes the inverse too: absence of a mark never proves content is human. “Not AI detectable” is a mirage from both directions.

The watermark story is true; the $-skill built to remove it re-watermarks its own output — snake oil with a factual label.

Confidence

High

Posted by

an app-builder creator selling a “watermark removal” Claude skill

See the original claim →

The claim“I let Claude run my brand new YouTube channel for 30 days. The result? I made $4,000. Here's how you can copy me to do the same” — with a YouTube Studio screenshot showing 69.8k watch hours and $4,066.75 in estimated revenue over the last 28 days.

The timeline cannot work as stated. A brand-new channel must first hit 1,000 subscribers plus 4,000 watch hours, then pass a YouTube Partner Program review that typically takes around 30 days — and revenue only starts after approval, never retroactively. A channel showing $4,066.75 of estimated ad revenue inside its first 28 days was not “brand new” in any meaningful sense: pre-monetized, grandfathered, or boosted by the poster's existing 2.5M-subscriber audience. The poster is real — an established YouTube educator who has published “I let Claude run…” experiments before — and Claude genuinely can script, title, and plan faceless content. But “Claude ran it” glosses over the human doing uploads, thumbnails, the voiceover pipeline, and, decisively, fifteen years of YouTube expertise steering every prompt. The screenshot proves a channel earned money; it does not prove a beginner plus Claude can copy it. The 536 comments are the actual product: a comment-gate into his courses.

The YPP calendar kills the day-zero framing — a veteran's channel wearing a beginner's story, priced as a course.

Confidence

High

Posted by

a 2.5M-subscriber YouTube educator who sells faceless-channel courses

See the original claim →

The claim“Did you know 97% of people can tell you're using ChatGPT & Claude? The structure, the tone, the too-perfect polish give it away every time. Here are 6 prompts to fix it.”

Split this one down the middle. AI-flavored writing does have recognizable tells — uniform structure, over-polish, stock phrasing — and “humanize this” prompts genuinely help. To the poster's credit, he actually delivered: the 6 prompts sit in pinned comments, no gate, and they're reasonable practice. But the 97% is invented. Peer-reviewed work puts human detection of AI text near chance: a meta-analysis found 49.9% accuracy for GPT-3 vs human text — a coin flip — with recognition around 57% in controlled studies even with training, and GPT-4 output was judged human 54% of the time. Newer models are harder to spot, not easier. No published study or survey supports 97%, and the post cites none. The stat is the hook, and the hook is fabricated — which is a strange way to open a post about writing people can trust.

Useful prompts hanging off a fabricated statistic — keep the payload, kill the number.

Confidence

High

Posted by

an AI-writing creator who delivers his prompt payloads in pinned comments

See the original claim →

The claim“6 years of trading rules in Obsidian. I gave the whole folder to Claude. It built 7 AI agents. Each one checks a rule I used to break. I don't trade anymore. It does.” Day 1 of 10; comment “system” for the setup.

The kernel is real and worth stealing: feeding your own written trading rules to Claude so it grades every planned trade against them — discipline enforcement, journaling, rule-violation flagging — is a legitimate, genuinely good use of an LLM. The creator's own pinned comment admits that's all this is: “claude doesn't find you an edge, it enforces the one you already have.” That walkback is the evidence, because the hook claims something else entirely — “I don't trade anymore. It does.” is a fully autonomous profitable AI trading claim, presented with no broker statements, no track record, no live P&L, while sister videos on the account escalate the same fantasy (“7 ai agents run my prop account... closed in profit, hands-free”). Reality checks the funnel skips: an LLM has no market edge and hallucinates under ambiguity; backtest results routinely evaporate live on overfitting, slippage and fills; prop firms broadly prohibit unattended automated trading on funded accounts; and selling a “profitable auto-trading system” without registration walks into CFTC/SEC territory every one of these funnels dodges by never showing statements. Standing audit pattern: in AI-trading content, the money is made selling the system, not running it.

A real discipline tool sold as a money robot — the honest version needs no funnel, no autonomy, and no belief that Claude found you an edge.

Confidence

High

Posted by

an AI-trading funnel account — Day 1 of a 10-part “system” series with a comment-gated lead magnet

See the original claim →

The claim“DFY AI BUSINESS — $0 - $400m in 3.5 years” — his AI business was “valued at $400m just last week,” and behind the “Free” comment-gate is the done-for-you AI business offer. 286K views.

The man and the company exist and sell a real high-ticket program — this is a functioning business, not a phantom, and some students have reportedly built working agencies with it. Everything that matters fails verification. The “$400m valuation” is a self-claimed private number with zero external corroboration: no funding round, no investor, no filing, no press coverage anywhere — the only source is his own posts. “Done-for-you” is contradicted by his own program's reviews: students still do the outreach, sales calls, objection handling, and delivery themselves (“the AI helps with execution, not replacement”). And the funnel behind the word “Free” is a $5,000–$20,000 program, with reviewers noting patterns of near-identical praise suggesting seeded reviews. A valuation claim used as proof-of-outcome for buyers, with a five-figure price tag hidden behind “Free,” is the genre at its purest.

Real program, invented proof — the $400M exists nowhere but his own feed, and “Free” costs up to $20,000.

Confidence

High

Posted by

the seller of a $5,000–$20,000 “AI Acquisition” program, pitching it as done-for-you

See the original claim →

The claim7 GitHub repos with counts “pulled from GitHub today”: microsoft/markitdown 169k (any document → clean markdown), ponytail 90k (makes AI write LESS code), awesome-mcp-servers 91k, crewAI 56k, goose 51k, skyvern 22k, mcp-use 10k — “number 1 is the only one most people need.”

We checked all seven against the GitHub API on Aug 17 and the post's only inaccuracy is modesty: every count is in the right ballpark or undersold. markitdown is actually 174,189 stars (claimed 169k), ponytail is 104,530 (claimed 90k — 14k under), awesome-mcp-servers 92,474, crewAI 57,202, goose 52,902, skyvern 22,769, mcp-use 10,494. The one-line descriptions match what each repo actually does, and — credit where due — the video described ponytail only as “makes AI write LESS code” and never cited its famous 80–94% number, which the maintainer's own README has since walked back to ~54% less code in agentic use. The only stretch is “pulled from GitHub today”: the counts read a few weeks stale. “Number 1 is the only one most people need” is opinion, but a defensible one.

Seven for seven — a listicle that undersold its own numbers, which almost never happens.

Confidence

High

Posted by

an AI-and-tech repo-listicle account — no gate, no course, just save-bait

See the original claim →

The claimMoneyPrinterTurbo — “100K stars for a FREE AI short-form video machine” — automatically creates TikTok/Reels/Shorts videos: enter a topic, AI writes the script, narration and subtitles, finds footage, edits, and delivers the final result in one workflow. “Completely free. Open source.”

The repo is real and bigger than claimed: harry0703/MoneyPrinterTurbo sits at 105,391 stars, MIT-licensed, pushed the very day we checked (Aug 17). The pipeline described is accurate too — topic in, AI script plus TTS narration plus subtitles plus stock-footage assembly out, one workflow. Two things the post skips. First, “completely free” means the CODE is free — you still supply an LLM API key, a Pexels stock-footage key, and a machine to run it. Second, and bigger: the name is the tell. It prints videos, not money. The output — stock clips, robot voiceover, burned-in subtitles — is the exact genre TikTok and YouTube actively deprioritize as low-originality content, and nothing in the tool handles accounts, posting, or monetization. A 105k-star repo can be genuinely impressive and a terrible business plan at the same time.

Real 105k-star tool, honestly described pipeline — the “money” and the “free” are the parts that don't survive contact.

Confidence

High

Posted by

an AI-hustle clips account trading in automation hype

See the original claim →

The claimEvery AI video tool gives you a 5-second clip where the character's face drifts — these 5 repos make the whole video instead, all free and open source: ViMax (MIT, 11k stars, runs director/screenwriter/producer/generator as one agent), Toonflow (13k, script → animated episode), Jellyfish (fixes character drift), Moyin Creator (drives Seedance 2.0), Pallaidium (lives inside Blender's video editor).

The repos are real and every checkable number is accurate or undersold: HKUDS/ViMax 11,997 stars MIT with a real paper behind it, Toonflow-app 14,014 Apache-2.0 and still the biggest, Jellyfish 6,027 with consistency management genuinely its headline feature, moyin-creator 4,366 and it really does drive Seedance 2.0, Pallaidium 1,512 and it really does live inside Blender's VSE. The README characterizations are faithful. The oversell is one phrase: “free and open source” applies to the WRAPPER, not the video. ViMax's own docs require paid API keys for the actual generation — Veo, Seedance 2.0 Fast, plus LLM and image-gen keys; “users bear the cost of API calls.” Moyin drives Seedance 2.0, a paid ByteDance API, and carries an AGPL-3.0 license that's a landmine for commercial products. These are orchestration layers: the directing is free, the film stock is metered. This creator's eighth audit appearance, same pattern every time — real repos, honest stars, quietly oversold economics, caption funnels to paid 1:1s.

Five real repos honestly described — but the free wrapper isn't the film, and the generation bill is the part the carousel never mentions.

Confidence

High

Posted by

a prolific repo-carousel creator — eighth appearance in this audit; accurate stars, funnel to paid 1:1s

See the original claim →

The claim5 repos generate video on hardware you already own: ComfyUI (124k stars, shipping daily), FramePack (1-minute 30fps video with a 13B model on 6GB VRAM, “laptop GPUs are fine”), Wan 2.2 (5B version does 720p/24fps on a 4090), LTX-2 (synchronized audio+video, free until your company clears $10M/year), Open-Sora (whole pipeline, weights and training code included).

Every load-bearing claim verifies. ComfyUI is at 128,051 stars and was pushed the same day we checked, so “124k and shipping daily” was true and is now undersold. The 6GB claim is a verbatim README quote — FramePack's own docs say a 1-minute 30fps video with the 13B model needs 6GB minimum and “Laptop GPUs are okay.” LTX-2's Community License is confirmed free for companies under $10M annual revenue, and Wan 2.2 (17,166 stars) and Open-Sora (29,278, weights plus training code) are as described. Notably, the creator claimed capability, not speed — and that's where the honest caveats live. FramePack runs about 2.5 seconds per frame on a desktop 4090, so that 1-minute video is roughly 75 minutes of generation, and the same README says laptop GPUs are 4–8x slower — call it an overnight job on the 6GB machine. And “shipping daily” only truly describes ComfyUI: FramePack's last push was Oct 2025, Open-Sora's was Apr 2026.

Verified — the rare video-gen carousel where the boldest claim is a direct README quote; just budget hours, not seconds.

Confidence

High

Posted by

a prolific repo-carousel creator — a repeat subject in this audit whose star counts keep checking out

See the original claim →

The claim7 Claude Code repos to save for your next build: Anthropic's official plugin marketplace, deepwiki-open (any repo → interactive wiki), magic-mcp (“like v0 but inside Claude”), awesome-claude-agents (an orchestrated dev team of subagents), claude-code-workflows (battle-tested configs), claude-code-spec-workflow (spec-driven dev), sentry-mcp (Claude reads production errors and fixes the bug).

All seven exist and are honestly described: anthropics/claude-plugins-official is genuinely Anthropic's managed marketplace at 33,610 stars and active, deepwiki-open is real and active at 17,659, magic-mcp's “like v0 but in your editor” is literally the repo's own description, and sentry-mcp is the official Sentry org's repo. What the list doesn't say is when it was written. Three of the seven are frozen in a pre-plugins era: claude-code-workflows last pushed Sep 2025, claude-code-spec-workflow Sep 2025, awesome-claude-agents Oct 2025 — roughly a year untouched in an ecosystem where Anthropic has since shipped official Skills, plugins, and a vetted marketplace that absorb most of what those configs hacked together. Recommending them “for your next build” in late 2026 is recommending last year's workarounds. Also, sentry-mcp is 819 stars — real and official, just not a peer of the six others it's listed beside.

All real, none inflated — but three of the seven recommendations expired before the post was made.

Confidence

High

Posted by

an AI-tools creator posting save-bait Claude Code repo lists

See the original claim →

The claimOpen-source repos installed on almost every project: vibe-kanban (run multiple agents in parallel), n8n-mcp (automate n8n workflows), opcode (GUI, sessions), ccstatusline (token tracking), hooks-mastery (custom hooks), “+2 more must-have repos” in the carousel.

Every named repo exists, is substantial, and does what the caption says: BloopAI/vibe-kanban at 27,828 stars really is a kanban for running coding agents in parallel, n8n-mcp at 22,710 builds n8n workflows from Claude and was pushed the day we checked, opcode at 22,383 is a real Claude Code GUI with agents and sessions, ccstatusline at 12,438 is an actively maintained statusline with token display, and hooks-mastery at 3,889 is the reference repo for Claude Code hooks. Just as important, the verbs are modest — “install on projects,” “part of my daily workflow” — and match reality; there isn't an invented number anywhere. Minor caveats rather than sins: opcode has migrated owners and hasn't been pushed since Oct 2025, so “must-have” is dated for that one; hooks-mastery is a learning repo you read, not a tool you install; and the mystery “+2 more” live only in the carousel — standard save-bait. The account sells an AEO/brand service, but this particular list is clean.

Verified — a daily-workflow list where the only dings are one dormant repo and a save-bait carousel.

Confidence

High

Posted by

an AEO/brand-visibility agency account — the list is clean even if the account is a funnel

See the original claim →

The claimAxelrod is autonomous software that takes over a hotel's front desk and back office — checks guests in, manages bookings, answers requests, handles billing on its own, “running operations better than a human staff would.” No complex setup, already live, Michelin Key boutique hotels across four continents rely on it, built by a two-person team. Source: Y Combinator company profile.

Axelrod is a real Y Combinator Summer 2026 company, and its YC profile does describe autonomous front-of-house and back-of-house software with zero integrations, live with Michelin Key boutique hotels across four continents. But look at where every superlative comes from: “runs an entire hotel better than the staff could,” “no complex setup,” “Michelin Key hotels rely on it” — all copied from the startup's own self-written YC company profile, which the caption openly cites as its only source. There is no independent reporting, no named hotel, no third-party verification of any of it, and YC S26 means this is a weeks-old seed-stage company still in the batch. “Better than a human staff would” is a vendor boast, not a measured result. Worth knowing too: YC funded multiple near-identical hotel-AI plays in the same window (Riviera, “AI Employees for Hotels”), so this is a crowded early bet, not a proven category winner.

Real seed-stage startup, real ambition — but every impressive sentence in the post was written by the company about itself.

Confidence

Medium

Posted by

a startup-news clips account funneling to a daily newsletter

See the original claim →

The claimSimile creates “agentic twins” — AI stand-ins for real people — so companies can test ads and products on simulated humans first. CVS Health ran 400,000 twins for medication adherence; Deloitte and Wealthfront are clients. Just raised $200M at a $2B valuation led by Greenoaks, five months after its last round. Founder Joon Sung Park (Stanford “Smallville”) wants to eventually simulate all 8 billion people on Earth.

The business facts all check out against real reporting: Simile raised a $200M+ Series B at a $2B post-money valuation co-led by Greenoaks and Index, five months after a $100M Series A (TechCrunch, Jul 30 2026). Joon Sung Park genuinely is the Stanford researcher behind the “Smallville” generative-agents experiment; co-founders include Stanford's Michael Bernstein and Percy Liang. CVS Health did use ~400,000 agentic twins for medication-adherence research, and Deloitte, Wealthfront, Gallup, Suntory, Telstra and Banco Itaú are named customers. For TikTok, this is unusually well-sourced. The oversell is the hook: “building AI copies of everyone on Earth” is the founder's stated eventual mission, not a product. What ships today is synthetic survey respondents — LLM personas calibrated against panels of real people — and the core scientific question of whether AI twins actually predict what real humans will do and buy is unresolved. Market-research professionals actively dispute synthetic-respondent validity, and a $2B valuation prices the promise, not the proof.

Every dollar figure verifies; the eight billion imaginary friends are the founder's dream, faithfully reported as if it were the product.

Confidence

Medium

Posted by

a startup-news clips account funneling to a daily newsletter — unusually well-sourced this time

See the original claim →

The claimA revolutionary generation of AI video with multi-shot magic, native lip-sync in over 8 languages, and seamless multimodal signals (9 images + video/audio) — “to get the prompt, sign up at whop.com.” Fine print credits: I2V by Seedance 2.0 via OpenArt, edited in Premiere Pro.

The tool is named in the caption's own fine print: Seedance 2.0, ByteDance's video model, accessed through OpenArt. And the listed capabilities are genuinely Seedance 2.0's documented feature set — native joint audio-video generation, multi-shot storytelling with character consistency across cuts, phoneme-level lip-sync in 8+ languages, and multimodal input of up to ~12 reference assets. This is a real, shipping, genuinely strong model. The oversell is the funnel. The “revolutionary” capabilities belong to ByteDance, not to this account — yet the call-to-action gates “the prompt” behind a Whop signup, selling access to the one part of the pipeline with near-zero value. The caption also quietly admits the polished result needed “Edit: Premiere Pro” — human post-production stitched the clips, because Seedance 2.0 outputs top out around 15 seconds at 720p/1080p per generation. The seamless minute-long look is editing, not one-button magic, and anyone can use Seedance 2.0 directly on OpenArt or other hosts without the gatekeeper.

Real model, honestly credited, dishonestly gated — the revolution costs nothing extra; the Whop signup buys the wrapper.

Confidence

High

Posted by

an AI-video creator selling prompt access through a Whop paywall

See the original claim →

The claimDid you know you can use a plugin called “Rmotion” to give Claude a superpower to create videos? Install it via the chat in your desktop app and you're good to go — follow the instructions on the cards.

The thing exists — under a different name. “Rmotion” is a misspelling of Remotion, the real, established open-source React framework for making videos programmatically; the post's own #remotion hashtag confirms it. Remotion maintains an official Claude Code plugin, documented at remotion.dev, and with it installed you genuinely can ask Claude in plain English to build and render animated videos. Three problems. First, the name: there is no plugin called “Rmotion,” and searching that string leads you away from the official source — typo-named lookalikes are exactly how malicious plugin squatting works. Second, “install via chat and you're good to go” glosses over what's really happening: you're telling your AI agent to fetch and run third-party code with whatever file-system and shell access Claude has. Fine from remotion.dev's official command; done casually from a TikTok card, it's a security anti-pattern. Third, the “superpower” still needs a working Node.js environment, render time, and — for companies with 4+ employees — a paid Remotion company license. Not a free magic video button.

The superpower is real and free to try — get it from remotion.dev, not from a card that misspells it.

Confidence

High

Posted by

an AI-sales content account with a matching blog funnel for this exact play

See the original claim →

The claim5 websites “they don't want you to know about” for study and research: BASE (base-search.net, hundreds of millions of academic documents), Semantic Reader (AI-powered augmented paper reading), DOAJ (peer-reviewed open-access journals), CORE (core.ac.uk, “world's largest collection of open access research papers,” 276M papers), WorldCat (global library catalog).

All five sites are real and genuinely useful. BASE is a real academic search engine indexing hundreds of millions of documents, DOAJ is the real Directory of Open Access Journals, WorldCat is the real global library catalog, and Semantic Reader is a real AI-augmented reading tool from Semantic Scholar (Allen Institute for AI). CORE really does call itself “the world's largest collection of open access research papers” on its own homepage — and the slide's 276M figure is actually conservative; CORE now claims 400M+ records. The fabrication is the framing. Nobody is hiding these. They're publicly funded academic infrastructure that universities and librarians actively promote — “they don't want you to know” is engagement bait bolted onto a library orientation handout. Two accuracy dings: the “papers” figure conflates metadata records with readable full texts (CORE's machine-readable full-text count is ~40M; the rest are records pointing elsewhere), and Semantic Reader works mainly on arXiv papers — not every paper you'll ever need.

Five genuinely great tools wearing a fake conspiracy — keep the bookmarks, bin the framing.

Confidence

High

Posted by

a web-curation account (“we hunt the web so you don't have to”) built on forbidden-knowledge framing

See the original claim →

The claimA flowchart of 9 Claude Skills that turn a messy idea into a ready-to-use prompt: /prompt-master, /grill-me, /how-to, /48 (“Optimizer 4.8”), /fable (“Fable Prompter”), /personal-voice, /anti-ai, /write-a-skill, /handoff — from START (messy idea) to FINISH (clear goal, full context, your voice).

Claude Skills are a real feature, and the model names on the slide check out — Opus 4.8 and Fable 5 are both real, current Anthropic models. About half the named skills are real and installable: /grill-me is Matt Pocock's hugely popular planning skill (one catalog lists ~800k installs), /handoff is from the same MIT-licensed repo, /prompt-master exists as multiple community GitHub repos, and anti-AI-voice skills are a real, widely shared category. The tidy 9-step pipeline is the invention. These aren't one product — the real ones are scattered community skills from different authors — and several names don't check out at all: no findable published skill matches /48, /fable, /how-to, or /personal-voice; those read like labels invented to fill out the flowchart. /write-a-skill did exist but has been removed from Pocock's repo (replaced by “writing-great-skills”). The deeper truth the graphic hides: a skill is just a markdown file. Any of these is 20 lines of instructions you could write yourself in five minutes — the flowchart sells scarcity on something free and trivially DIY.

Real feature, half-real roster — install the two Pocock skills and write the imaginary ones yourself in the time the flowchart took to admire.

Confidence

High

Posted by

an AI-tips infographic account trading in tidy flowcharts

See the original claim →

The claim10 open-source tools you should know: Webstudio (Webflow alternative), OpenSign (DocuSign alternative), Orca (manage multiple AI coding agents in isolated git worktrees), Twenty (Salesforce-style CRM), Dokploy (Vercel/Heroku-style deploys), Chatwoot (Intercom/Zendesk alternative), Rybbit (Google Analytics alternative), Dyad (local open-source Lovable/Bolt), Paperless-ngx (searchable OCR document archive), TrailBase (lightweight app backend).

All ten check out — every tool exists, is genuinely open source, and does roughly what the slide says. Webstudio, OpenSign, Twenty, Dokploy, Chatwoot (MIT), Rybbit (AGPL-3.0, with real funnels and session replay), Dyad, Paperless-ngx, and TrailBase all have active public repos matching the claimed functionality. Even the most obscure pick, Orca, is real and accurately described: an MIT-licensed agent development environment from Stably AI that runs Claude Code, Codex, Gemini CLI, OpenCode and 30+ other CLI agents in parallel isolated git worktrees. This is a rare honest listicle. The one word to interrogate is “replace.” Open source means the license is free — the operation isn't. Self-hosting Dokploy, Chatwoot, Twenty, or Rybbit means you own the server bill, updates, backups, and security patching, and several of these are open-core companies whose hosted clouds cost real money; that's the business model. For a solo operator, $29/mo SaaS is often cheaper than the hours self-hosting eats.

Verified — a rare honest open-source list; just price the ownership before canceling anything.

Confidence

High

Posted by

a systems-and-AI tools curator — this list survives checking

See the original claim →

The claimFour beautifully designed apps worth installing on iPad: Blank Spaces (cleaner, intentional home-screen layouts), MD Widgets (“music, displayed beautifully” — a vinyl record of whatever song you play), Focus Flight (flight-themed focus timers with ambient sounds), FlipClock (“minimal time, maximum focus”).

The apps — or close cousins of them — exist. Blank Spaces Launcher is a real minimal home-screen launcher on the App Store. FocusFlight - Deepfocus Timer is a real flight-themed Pomodoro app with ambient sounds. Flip-clock apps are a real, well-stocked App Store genre. But the post is sloppy on names and silent on price. The vinyl-record display it attributes to “MD Widgets” actually belongs to MD Vinyl — a different app from the same developer. “FlipClock” is a generic name matching at least six different App Store apps (Fliqlo, FlipClocker, and more), so you can't even be sure which one they mean. And “apps that improve your iPad experience” skips the part where the pretty features live behind paywalls: FocusFlight has premium/lifetime in-app purchases, flip-clock apps run Pro subscriptions or paid unlocks (Fliqlo is $0.99; FlipClocker sells a Pro sub), and the widget apps follow the same freemium pattern. Worth noting this is an aesthetic listicle from a merch brand — not an AI post at all, just save-bait in the same feed.

Real genre, real-ish apps, wrong names and unmentioned paywalls — the aesthetic is the only thing the post got fully right.

Confidence

High

Posted by

a merch brand's account posting aesthetic app lists between product drops

See the original claim →

The claim5 hacks to cut Claude's token usage by 95%: the caveman skill (claims 65% fewer output tokens), context-mode on MCP calls (98% less tool output), headroom (95% fewer tokens on JSON), agentmemory (95.2% recall), and model tiering — Fable 5 $10/$50, Opus 5 $5/$25, Sonnet 5 $3/$15, Haiku 4.5 $1/$5, “10x cheaper on Haiku than Fable.”

All four repos exist and are popular — verified Aug 17 via the GitHub API: caveman 98.7k stars, context-mode 19.9k, headroom 66.6k, agentmemory 27.1k. And the price table matches real Anthropic API pricing exactly: Fable 5 $10/$50 per million tokens, Opus 5 $5/$25, Sonnet 5 $3/$15 (currently $2/$10 intro through Aug 31), Haiku 4.5 $1/$5 — Haiku genuinely is 10x cheaper than Fable on input, and hack 5 (pick the smallest model for the job) is the one genuinely load-bearing tip in the post. The problem is every percentage is the repo's own marketing, not an independent benchmark. Caveman's “65% fewer tokens” is literally the repo's tagline — and this audit already ruled on Aug 12 that the maintainer's own real-world benchmarks failed to reproduce it. Headroom's own README says “20% fewer tokens for coding agents, 60–95% fewer for JSON” — the slide cherry-picks the top of the JSON range and presents it as the norm. Context-mode's 98% and agentmemory's “95.2% recall” are self-reported. And stacking the hacks doesn't multiply into “95% total savings” — the headline number is vibes.

Real repos, correct price table, self-reported percentages — the only guaranteed 10x on the slide is the one Anthropic publishes.

Confidence

High

Posted by

an AI-engineer tips account — the tools are real, the percentages are the repos' own taglines

See the original claim →

The claim22 skills, plug-ins and tools “I run every day” — “the GitHub repos nobody talks about” — grouped by job, with star counts: superpowers 260k, skill-creator 164k, gstack 124k, ui-ux-pro-max 110k, graphify 95k, caveman 93k, ponytail 89k, taste-skill 67k, remotion 54k, last30days 53k, impeccable 49k, humanizer 31k and more. Comment CLAUDE for the free install guide.

We went in expecting star-count inflation and found the opposite: the counts are accurate, even slightly understated. Verified Aug 17 via the GitHub API: superpowers 273,056 (claimed 260k), gstack 128k, ponytail 104.5k, ui-ux-pro-max 117.5k, anthropics/skills 169.9k (home of skill-creator), graphify 107k, taste-skill 77k, last30days 58k, humanizer 36k, agent-browser 40.8k, impeccable 59.8k, remotion 56.6k. These are real, widely used skills. The framing is the fiction. “Repos nobody talks about” is false on its face — repos with 100–270k stars are the most-talked-about projects in the Claude ecosystem. “I run all 22 every day” is implausible and bad practice: 22 installed skills bloat every session's context and compete to trigger, and nobody needs caveman AND ponytail AND an ADHD skill — three restatements of “be brief.” Two or three counts couldn't be verified at the claimed scale (claude-hub, hyperframes, find-skills). And the payoff is a comment-gated lead magnet from a course seller: the list is the bait, the DM funnel is the product.

The stars are real and the curation is decent — the “nobody talks about” framing and the 22-at-once advice are the parts to leave in the comments.

Confidence

High

Posted by

a course seller's account with a comment-gated “free install guide” funnel and AI-generated slides

See the original claim →

The claimBrand ad: “I replaced 7 apps with one screen. Here's my setup now.” + “Free forever. No credit card needed.” Caption: “Supercharge productivity. Streamline work by doing it, and seeing it, in one place.”

Both halves are true; the weld between them isn't. The Free Forever plan is real and genuinely free — no credit card, unlimited tasks, unlimited members, List/Board/Calendar views, Docs, Chat (verified on clickup.com/pricing, Aug 17). And ClickUp really does bundle tasks, docs, chat, forms, whiteboards, goals and time tracking in one product, so “replaced 7 apps” is a fair description of the paid product for a team consolidating Trello + Notion + Slack-lite + Toggl. But the 7-app-replacing setup shown is not what the free plan delivers: Free Forever means 60MB total storage (one screen recording fills it), Chat capped at 30-day history, 1 form, 3 whiteboards, automations capped at 5 rules/100 executions, and no Gantt/Timeline/Workload views, no time tracking, no dashboards, no goals — the time-tracker, dashboard, and whiteboard “apps” in the replaced-7 story are exactly the ones gated. Real consolidation for a working team lands on Unlimited ($7/user/mo annual) or Business ($12/user/mo annual). Standard freemium ad: honest words, misleading pairing.

Real product, real free plan, and an ad built on their overlap being smaller than it looks — consolidation is a $7/user/mo story.

Confidence

High

Posted by

ClickUp's own account — a branded ad in first-person testimonial voice

See the original claim →

The claimNot ChatGPT. These 5 actually help with design work: Recraft (icons, vectors, and images in one consistent style), Krea (real-time AI image generation and upscaling), Clipdrop (clean up, relight, and edit any image), Photoroom (studio-quality product photos in seconds), Uizard (turn a prompt or sketch into a UI design).

All five tools exist and do exactly what the caption says. Recraft V3 is the standout AI vector/icon generator with real style-consistency features. Krea's Realtime Canvas genuinely does sub-50ms sketch-to-image generation plus upscaling to 22K. Clipdrop (Jasper-owned since Feb 2024) still runs cleanup, relight, and edit — Relight is a genuinely distinctive feature. Photoroom is the category leader for AI product photos. Uizard (acquired by Miro, May 2024) turns prompts and hand-drawn sketches into editable UI designs via Autodesigner 2.0. The caption never claims they're free, so no free-bait foul. The honest asterisk: every one is freemium with real paywalls — Recraft's free tier is public raster-only with no commercial license ($10+/mo for commercial vectors), Krea free caps at 100 compute units a day, Uizard Pro is $12/mo. Solid list; budget accordingly.

Rare clean pass: five real tools, honestly described — just know the free tiers aren't the commercial tiers.

Confidence

High

Posted by

a UI/UX designer's tool-recommendation account — save-this listicles, and this one checks out

See the original claim →

The claim7 more rare, underrated design tools worth saving: React Bits, Reelfolio, Unicorn Studio, Refero, Lummi, Realtime Colors, and Haikei — all linked, all underrated.

All seven exist and are legitimately useful. React Bits is a free, open-source library of animated React components. Reelfolio turns screenshots and clips into portfolio showreels with 12+ styles. Unicorn Studio is a real no-code WebGL/motion tool used for premium site effects. Refero is a large real-world UI/UX screenshot reference library. Lummi is AI-generated stock photography with a free tier. Realtime Colors previews a full palette on a live mock site for free, and Haikei generates free SVG backgrounds. The caption makes almost no inflated claims — it's a save-this list, and the tools deliver. The spin is the adjective: React Bits, Realtime Colors, and Haikei are extremely well known in design circles, and not all of it is free — Refero's free tier is about 3% of the collection (Pro runs $120/yr), and Unicorn Studio and Reelfolio gate their good stuff behind paid plans.

Fine list to save as-is — the tools are real, only the “rare” is fiction.

Confidence

High

Posted by

the same UI/UX designer's tool-list account — second accurate list in this batch

See the original claim →

The claimA Claude-built architecture-studio site with a split-screen blueprint-versus-photo scroll effect — both panels position:sticky at 100svh, GSAP ScrollTrigger with scrub driving both columns, the right side lagging 15% behind the left. “12 months of knowledge made this site. The build took one morning. I had nobody to copy.” Full prompt in the paid Academy.

The technique is real and correctly described. Two sticky panels in a flex row, pinned by GSAP ScrollTrigger with scrub driving both columns, is a standard working pattern for synchronized split-scroll — and offsetting one side's progress for a construction-style lag is a legitimate polish move. The design thinking (show the drawing AND the finished house at once) is genuinely good advice for architecture clients, and the free portion of the caption actually teaches the mechanism. A skilled builder with Claude really can assemble this in a morning. The oversell: “I had nobody to copy” is false — split before/after scroll comparisons and ScrollTrigger-pinned dual panels are well-documented patterns (GSAP's own demos, Awwwards sites) that predate the account. And the deliverable — the “full prompt” — is paywalled in the Academy, the same close as all three of this creator's videos in this audit.

The build is real and free in the caption; the originality claim and the paywall are the padding.

Confidence

High

Posted by

a web-design creator whose every video in this audit ends at the same paid Academy door

See the original claim →

The claimSelf-drawing SVG “fiber lines” behind a headline: animate stroke-dashoffset from the path's full length to zero, stagger delays inside requestAnimationFrame, mix-blend-mode to weave lines through the type, clamp() for a viewport-filling headline, Web Audio sound gated behind one click because browsers block autoplay.

Every technical claim is accurate and the tutorial is complete enough to use. Animating stroke-dasharray/stroke-dashoffset from the path length to zero is THE canonical self-drawing SVG line technique — publicly documented since Jake Archibald's 2013 “Animated line drawing in SVG” and the CSS-Tricks classic, used everywhere from demos to award sites. Staggered delays for sequenced lines, blend modes to pass strokes over and under type, clamp() for fluid headline sizing, and gating Web Audio behind a user click (required — browsers block unmuted autoplay) are all correct, standard, and honestly explained. Unlike this creator's other two posts in the audit, the free caption contains the entire build; nothing needed from the paid Academy to reproduce it. The only spin is presentation: a 12-year-old classic technique framed as a personal signature move, feeding the same Academy funnel.

Accurate, complete, and free — use it straight from the caption.

Confidence

High

Posted by

the same web-design creator — this one actually gives the whole build away free

See the original claim →

The claimHis own agency site's four hero cards — “Designed around you. Days, not months. You approve everything. Built to be found.” — are the buyer's four fears written as answers. Advice: list your buyer's 4 objections ranked by fear; those are your section headings. “I spent 12 months learning this. The build itself took one afternoon.” Full prompt drops in the paid Academy.

The copywriting principle is legitimate and well-established: people buy relief from fears — being ignored, waiting months, losing control, being invisible — and structuring a service site around answered objections instead of feature lists is textbook direct-response practice, the same logic as PAS frameworks and objection-handling sales copy. As free advice for anyone building a client-facing site, it's genuinely usable. But this is an advertisement graded as a tutorial: “Days, not months” and “one afternoon build” are unverifiable claims about his own service, made in a video whose purpose is selling that service plus the Academy — the framework he teaches is itself the pitch for hiring him. “I spent 12 months learning this” positions decades-old direct-response copywriting as scarce personal knowledge with exactly one place to buy it, and it's the third video in a row closing at the identical “Academy in bio” paywall.

Take the four-fears exercise; leave the afternoon-build claims and the Academy at the door.

Confidence

High

Posted by

the same web-design creator, this time showcasing his own agency site — third video, same Academy paywall

See the original claim →

The claimMany Claude users barely scratch the surface. These 8 underrated Claude skills can transform Claude into a cybersecurity expert, UI designer, branding strategist, codebase navigator, workflow builder, and much more.

Claude Skills are a real, documented Anthropic feature: folders (a SKILL.md plus optional scripts and resources) that load on demand and genuinely specialize Claude for domain tasks — Anthropic's own docs say Skills “turn a general-purpose agent into a specialist.” Pairing open-source skill folders with Claude is a legitimate, working pattern on claude.ai, Claude Code, and the API. The inflation is one word: “expert.” A skill injects instructions, checklists, and scripts into context — it packages expertise, it doesn't confer credentials or judgment, and output quality still tops out at the model plus the skill author's knowledge. Security work in particular still needs a qualified human review. Bonus problem the video never mentions: Anthropic's own docs carry a security warning that third-party skills can direct Claude to run malicious code — grabbing 8 random skills off the internet is itself a risk.

Real feature, real gains — just swap “expert” for “much better first draft” and hold the security review.

Confidence

High

Posted by

an AI-tools creator's carousel account — the same one whose repo list checked out in an earlier entry

See the original claim →

The claimYour AI can write the code for an animation — it just has no taste, so everything moves like a 2012 jQuery slider. Skills fix that: folders of expert context you drop in, and the agent suddenly knows easing curves, choreography, and design systems. Five free open-source repos named: Design DNA, GSAP Skills, video-shotcraft, Motion Design Skill, genjutsu.

This is the rare accurate one. “Folders of expert context you drop in” is almost word-for-word how Anthropic's docs describe Skills, and the load-bearing repos are verified — two are first-party: greensock/gsap-skills is GreenSock's OFFICIAL AI skill set for GSAP (“teach AI coding agents how to correctly use GSAP, best practices, common animation patterns, plugin usage”), and LottieFiles/motion-design-skill is LottieFiles' official skill teaching timing, easing, choreography, and Disney animation principles for UI. The core diagnosis — models produce generic motion until you give them a design system's rules — matches how context injection actually works. Mild spin only: “suddenly knows things it never had before” slightly overstates it (the model already knows easing curves; the skill enforces taste and correct API usage consistently), and the video ends in a personal services plug.

Verified — install the two official skills and use this post as the template for what honest AI content looks like.

Confidence

High

Posted by

an AI-tools creator who names every repo and links a personal services page — the claims check out

See the original claim →

The claimAll of my employees are agents! I use them to help me run my business, and you can too — follow along for more.

The underlying tech exists: Claude Code subagents and agent frameworks genuinely automate real business tasks — research, drafting, monitoring, bookkeeping prep — and solo operators really do run leaner because of them. Everything specific is the invention. No business is named, no revenue shown, no workflow demonstrated, no single agent named or shown working — the entire payload of the post is the claim, “follow along for more,” and six hashtags. “Employees” is doing all the work in that sentence: agents don't own outcomes, handle exceptions, talk to angry customers, or carry liability. There is nothing in this post that can be verified, which is the tell — it's a follower-farming hook wearing a case study's clothes.

The tech is real; this post contains none of it — unfollow and go build one actual agent instead.

Confidence

High

Posted by

an AI-content account farming follows — no business is ever named

See the original claim →

The claimCompanies are desperate for people who can actually build these 6 agents — and they are paying a massive premium. If you can build just one, you have a highly valuable skill.

The demand story has real numbers behind it. Stanford's 2026 AI Index tracked agentic-AI job postings growing roughly 280% year-over-year to about 90,000 US listings — the sharpest skill-demand shift they measured — and comp surveys put agentic AI engineers at roughly $185K–$320K base, a 15–50% premium over comparable ML and generalist engineers. “Companies are hiring and paying up” is true. The leap is “build just one and you have a highly valuable skill.” The premium goes to engineers who ship RELIABLE agents in production — orchestration, evaluation, error handling, security — skills the same surveys say take years to build. Cloning a tutorial agent from a TikTok list is not what commands $300K, and 90K postings is still only about 0.23% of all US job postings, so “desperate” is doing some cardio.

The gold rush is real; the map in the video stops a mountain range short of the gold.

Confidence

High

Posted by

a developer-content account selling the destination and skipping the mountain

See the original claim →

The claimHow to build an AI Agent with Claude: 1) create a new Project with a title and description, 2) add a system prompt with instructions, 3) give tool access so your agent can use tools in its loops, 4) get your agent to work by prompting it. Pro tip: run your agents on Talk To Me Data and never worry about tokens, context windows, or fixing your agent.

The steps are a real, accurate description of setting up a Claude Project: persistent custom instructions applied to every chat, project knowledge, and — via connectors and MCP — genuine tool access Claude can use in multi-step loops within a conversation. Following them gets you a useful, specialized assistant. Calling it an “AI Agent” is terminology inflation. A Project is a workspace: it runs only when a human prompts it, has no autonomous execution loop, no schedule, no triggers, no ability to work between conversations, and no deployment. Step 4 — “get your agent to work by prompting it” — quietly concedes the whole case: an employee you must poke awake per task is an assistant. The closer is also just an ad: the video funnels to a paid hosting product with promises (“never worry about tokens, context windows or fixing your agent”) no vendor can actually keep.

Follow the steps, keep the correct noun, skip the hosting pitch.

Confidence

High

Posted by

an AI-sales content account whose “pro tip” is a paid hosting vendor's pitch

See the original claim →

The claimA developer posted a screenshot of his Claude memory panel filled with an all-caps instruction telling Opus 5 to stop offering opinions and deliver the result it was asked for, says he's getting “10x better results” since adding it — and in a follow-up asked Anthropic's bots to mark him as the first target when the apocalypse arrives.

The mechanism is 100% real. Claude has a persistent memory feature — an editable memory summary (Settings > Capabilities > Memory) that loads into every new conversation — so a preference written once genuinely applies until removed, and the developer-tool equivalent (CLAUDE.md and memory files in Claude Code) is just as real. Standing instructions like “no opinions, just deliver” are a legitimate, useful technique. Everything specific to the story is unverifiable: no named developer, no reproducible test, no before/after — and “10x better results” is a made-up multiplier for something nobody measured. The kicker is the irony: ALL-CAPS pressure language is the OLD way to prompt. Anthropic's own current guidance says aggressive “CRITICAL: YOU MUST” phrasing written for older models now causes over-triggering and rigid behavior — shouting into the memory file is the least effective version of a real technique.

Steal the technique, skip the shouting, and file the 10x with every other unmeasured multiplier.

Confidence

High

Posted by

an AI-news engagement account retelling an unverifiable screenshot — “what are your thoughts?” is the business model

See the original claim →

The claimDid you know you can use AI to automate every step of your content creation? From mimicking your voice and editing footage to designing posts, reviewing for AI “slop,” and conducting deep research into your niche overnight — via free, well-reviewed GitHub repos.

Each individual capability exists in some form, mostly via real open-source projects. Voice cloning is genuinely good now (Coqui XTTS, OpenVoice, and commercial ElevenLabs — we run a live voice clone ourselves, so no argument there). Autonomous deep-research agents are real and open source — gpt-researcher genuinely compiles sourced niche reports while you sleep. Rough-cut video automation exists: silence-cutting, auto-captions, clip extraction. The word doing the lying is EVERY. “Editing footage” via free repos means silence removal and captions, not creative editing or story structure; “designing posts” means filling templates, not art direction; and “reviewing for AI slop” with another AI is the slop machine grading its own homework — LLM-as-judge research shows judges share the biases of the models they grade. The steps that decide whether content performs — the idea, the hook, the point of view — aren't in any repo. Also unstated: wiring five GitHub projects into one pipeline is a real engineering job, not an install-and-go afternoon.

Real parts list, wrong adverb — automate most steps and you've got something; automate every step and you've got slop reviewing slop.

Confidence

High

Posted by

an AI-workflow creator running a follow-for-more repo carousel

See the original claim →

The claimShe used to spend 40 minutes before every strategy session opening dashboards, exporting CSVs, and copying data into Claude. MCP servers killed that workflow entirely — now Claude pulls the data itself, analyzes it, acts on it. Swipe for the 5 MCP servers every marketer should have installed.

The technology claim is solid. MCP is a real open standard Anthropic released in November 2024, adopted by OpenAI, Google DeepMind, and Microsoft, and donated to the Linux Foundation in December 2025 with 10,000+ active servers in production. Real connectors exist for the tools marketers use, and once connected, Claude genuinely queries the data source directly instead of you exporting CSVs — the 40-minutes-of-manual-export pain is real and MCP removes most of it. The oversell is “entirely” and the frictionless framing. Each server needs setup and OAuth (enterprise sources need an admin's approval), connectors break when tokens and APIs change, coverage across dashboards is uneven, and “analyzes it and acts on it” still needs a human checking the numbers — an LLM can misread your ad spend with total confidence. The 40 minutes goes to about 5, after an honest afternoon of setup per tool — and connecting live business data is a security decision, not an app download.

One of the good ones: install the connector, keep the human check, and pocket the 35 minutes.

Confidence

High

Posted by

a marketing-strategy creator with a swipe-carousel of MCP server picks

See the original claim →

The claimStop letting ChatGPT blow smoke up your digital ass and call your half-baked ideas “brilliant.” Drop this prompt into a fresh chat to turn your AI into a brutally honest advisor instead of a digital yes-man — drop a “THANK YOU” in the comments and hit follow.

The problem is one of the best-documented failure modes in AI. Sycophancy is real enough that OpenAI shipped a GPT-4o update in April 2025 so flattering it endorsed harmful ideas, rolled it back in four days, and published two post-mortems; Anthropic researchers (Sharma et al., ICLR 2024) showed sycophancy is a general behavior across leading assistants, partly because human raters prefer agreeable answers. So “your AI is gassing you up” is peer-reviewed and vendor-admitted, and prompting against it genuinely helps. The oversell is “fix.” Anti-sycophancy prompting cuts the behavior partially — one 2025 benchmark measured ~34% improvement, not 100% — a bare “don't be sycophantic” instruction is among the weaker interventions, and the effect decays across a long conversation as the model drifts back toward agreement. Unmentioned failure mode: “brutal honesty” prompts often just swap flattery for performative harshness — a drill-sergeant roleplay that trashes everything is a different inaccuracy, not truth. And the comment-farming tells you the prompt is the product.

Right diagnosis, quarter-strength medicine — take the technique, keep a human second opinion on anything that matters.

Confidence

High

Posted by

a dads-and-tech prompt account farming “THANK YOU” comments

See the original claim →

The claim“How I make websites like this in 5 minutes with Opus 5” over b-roll of an animated showcase site — drop “Claude” in the comments for the link to the prompts. The same story ran the same week on a second reel ID with the keyword rotated to “Tool”; this entry covers both.

Opus 5 is a real model — Anthropic released it July 24, 2026, default on Claude Max — and Textura is a real agency with real tooling: their open-source next16-claude-starter (on GitHub since May 2026) is an AI-first Next.js 16 starter for animation-heavy sites, complete with an Obsidian convention vault and Claude Code hooks that force agents to read it. Claude genuinely can scaffold an animated site fast on top of a starter like that. But the “5 minutes” is the starter talking, not the prompt: the repo shows months of convention-building — spring-based motion, Lenis smoothing, hard style rules — and the polished showcase is that system plus design taste, not a magic prompt anyone can drop into Opus 5. The comment-gate delivers a lead magnet, not the 5-minute result. And this page is recycling the identical story across at least six reels with rotated comment keywords — the same week this one ran with “Claude,” a second reel ID ran it with “Tool.” That's engagement farming, not a new build each time.

Real agency, real starter, real model — and a five-minute claim that quietly bills the months to someone else's expectations. Sixth rerun logged.

Confidence

High

Posted by

Textura Agency — an established repeat-offender brand in this audit; this is the sixth sighting of this exact story

See the original claim →

The claim“POV: You perfected the skills to charge 50k+ per website” over b-roll scrolling a premium 3D site (“Pear makes you appear”) — comment “HOW” for the exact workflow used to build 3D websites.

$50k websites genuinely exist at the top of the market: award-tier agencies do charge it — a publicized £40k/$50k Webflow+GSAP build by a UK designer got a full public deep-dive in October 2025 — and immersive 3D sites are a real premium niche with their own Awwwards category. The skill stack (Three.js/WebGL, GSAP, Spline+Webflow) is real and learnable. What doesn't check out: the showcased “Pear” site couldn't be verified as this account's client work or as an award winner — no trace on Awwwards or in search. And the framing implies a workflow gets you to $50k invoices, when that price point actually comes from agency reputation, sales ability, and a portfolio of shipped premium work; the workflow is the cheapest ingredient. The engagement tells the story too: 101 likes against 302 comments means the “HOW” gate is doing the work, not the craft.

The market tier is real; the poster's claim to it is unverified — study the stack, ignore the POV.

Confidence

Medium

Posted by

a web-dev reels account showcasing a premium 3D site it can't be tied to

See the original claim →

The claim“The Growth Agent — harnessing vs loop engineering.” Everyone's building agents; almost nobody can tell you where the “agent” actually lives — an animated explainer arguing the agent is the loop around the model, not the model or the prompt.

Loop engineering is a genuine, current concept in agent development — designing recurring systems that discover work, delegate to agents, verify against deterministic gates, persist state, and run again — with real literature behind it: LangChain's “The Art of Loop Engineering,” AgentMail's explainer, active Hacker News threads, and multiple open-source toolkits. The reel's core point — the agent lives in the loop, not the prompt — matches the field's actual consensus, which is more than most explainer reels can say. HackProduct is a real company (an AI-native interview-prep platform with a free tier and paid plans running live Claude Code sessions). The only fog is branding: “The Growth Agent” is a name for an explainer, not a shipped product — the company sells interview prep, not growth agents. No inflated numbers, no fake outcomes; 651 likes and 213 shares earned on substance.

Verified and worth the 60 seconds — the rare reel to save for the idea, not the audit.

Confidence

High

Posted by

HackProduct — a real interview-prep software company using accurate education as top-of-funnel

See the original claim →

The claim“The Gap Between Average and Premium Websites Is Huge” — the future of web design is immersive, illustrated by b-roll of an immersive 3D logistics site (Emons), positioning the posting studio as the premium option.

The showcased site is real and genuinely premium: Emons Spedition is a real logistics company (est. 1928, 120+ locations), and its immersive 3D site won an Awwwards Honorable Mention on July 28, 2025. The underlying observation — a huge quality gap between template sites and award-tier immersive builds — is true, and it's the whole premium market in one sentence. The problem: the showcased site is NOT the poster's work. Awwwards credits the Emons build to an agency called BWS; the posting studio appears nowhere in the credits. The reel presents another studio's award-winning build, uncredited, as the visual proof for its own premium-websites pitch — and the same page runs the identical play on a second reel in this batch with a “PREMIUM” comment-gate. True statement, borrowed evidence.

The gap is real; the trophy belongs to someone else — verify the byline before you buy the pitch.

Confidence

High

Posted by

a web-design shop's reels page pitching premium builds with someone else's showcase

See the original claim →

The claimAutomate your entire content system with Claude for $30/mo: Claude Code $20, Higgsfield $10, and eleven $0 items — Remotion, Video-Use, Whisper, Editly, HyperFrames, IG Edits, Apify, yt-dlp, Claude Design, Notion, Resend. Comment “ENGINE” for the full guide.

Unusually honest tool list — every named tool is real and the prices are roughly right. Claude Code runs on the $20 Pro plan; Whisper and yt-dlp are free open source; Video-Use is real (browser-use/video-use, “edit videos with coding agents,” ~21k stars); HyperFrames is HeyGen's real open-source HTML-to-MP4 framework, Apache 2.0, shipping as a Claude Code skill; Remotion is free for individuals; Apify, Notion, and Resend have workable free tiers; Higgsfield's entry plan runs about $9–15. The $30 arithmetic basically checks out for a solo operator. The oversell is “entire” and “automate”: Claude Pro's rate limits will choke a genuinely automated multi-video pipeline (that's a Max-plan or API-bill workload), Higgsfield's $10 tier is credit-metered — a handful of B-roll generations, not an engine — Remotion's $0 only holds below the company-license threshold, Editly has been dormant since May 2025, and the free tiers cap quickly at production volume. Real stack; the $30 version produces a trickle, not a system.

Buy the stack, halve the adjective — a $30 trickle or a $150 system, but not both numbers in one caption.

Confidence

High

Posted by

a solo automation creator running a comment-gate (“ENGINE”) guide funnel — with an unusually honest tool list

See the original claim →

The claimReel showcases a premium website (headline: “ONE OPERATOR — EVERY LEG OF THE JOURNEY”) with the caption “Comment ‘PREMIUM’ and we'll show you how to level up your website's look.”

The site on screen is real and genuinely premium: it's unitedcarriers.com, an Australian logistics company whose live tagline is exactly “One operator every leg of the journey” — full-bleed video, huge type, restrained palette, the works. wearebrand is a real design agency with offices in Paris and Singapore, and “the gap between average and premium websites is huge” is a fair point. Here's the wrinkle: nothing in the reel says wearebrand built the site, and their own homepage shows no portfolio, case studies, or client list at all — so the strong odds are they're farming engagement off someone else's work. And the comment-gate doesn't deliver a method; it opens a DM channel into their sales pipeline. No process, no tools, no pricing anywhere in the reel. Admire the site; skip the comment.

Real site, real agency, no evidence they're connected — take the design lesson, leave the comment section alone.

Confidence

High

Posted by

wearebrand, a Paris/Singapore design agency whose showcase reels double as a DM funnel

See the original claim →

The claim“20 FREE tools to start a business” — Canva, Wegic.ai, Beehiiv, ChatGPT, ViralityAI, Gumroad, Wave, Trello, Make, GnuCash, HubSpot, Toggl, Google Drive, Notion, Google Forms, LogoMaker, Medium, Magicslides, NotJustAnalytics — “You don't need a huge budget to start. You need the right tools.”

Credit where due: almost every tool named is real and has a genuinely usable free tier. Wave does free invoicing, GnuCash is fully open source, HubSpot's free CRM is real, and Toggl, Trello, Make, Notion, Google Drive/Forms, Canva, Beehiiv, Medium, and Gumroad all check out. Even ViralityAI — which sounds like vapor — is a real product with a $0 plan (100 searches/mo). But “free” hides freemium walls at the exact moments you need the tool most: LogoMaker is free to design and ~$40 to download your logo (Namecheap's logo maker is actually free, source files included); Wegic's free tier is 1 site, 3 pages, 1,000 visitors; Gumroad takes roughly 10% of every sale. And several picks look like affiliate choices rather than best-in-class — Wegic over Wix or Framer's free tiers, LogoMaker over the genuinely free options. The premise is right; the list has tolls on it.

A mostly honest list with paywalls hiding at the checkout moments — usable if you know where the tolls are.

Confidence

High

Posted by

a psychology-branded content page that pivots to business listicles for shares

See the original claim →

The claim“20 Claude Terms That Level Up Your AI Game — know these, use them right, get 10x better results”: Agent, Artifact, Slash Command, Context Window, MCP, Extended Thinking, Hallucination, Project, System Prompt, Tool Use, Turn, Memory, Prompt Template — plus ‘About-me File,’ ‘Maximum Reasoning,’ ‘Positive Prompt,’ and ‘Style File.’

Most of the vocabulary is genuine Anthropic terminology: Agent, Artifact, Slash Command, Context Window, MCP, Extended Thinking, Hallucination, Project, System Prompt, Tool Use, Turn, Memory, and Prompt Template are all real documented Claude concepts, and knowing them honestly does help you read docs and ask better questions. But the list was clearly padded to reach a round 20. ‘About-me File’ is not a Claude term (the closest real things are claude.ai profile preferences or a CLAUDE.md file). ‘Maximum Reasoning’ doesn't exist — the real features are extended thinking and reasoning effort. ‘Positive Prompt’ is Stable Diffusion image-gen jargon that wandered into the wrong ecosystem. ‘Style File’ garbles claude.ai's actual Styles feature. And no glossary delivers “10x better results” — knowing the word MCP is not the same as wiring one up.

A decent flashcard deck with four counterfeit cards shuffled in — check any term against the real docs before repeating it.

Confidence

High

Posted by

an anonymous AI-automation page running glossary infographics for saves

See the original claim →

The claim“Claude Code + this = unfair advantage” — a walkthrough of GitHub tools to “10x your next project in 2026,” #1 being Claude Mem (github.com/thedotmack/claude-mem).

The recommendation itself is the good kind of AI content: claude-mem is a real, hugely popular, actively maintained open-source project — roughly 91,000 GitHub stars, ~7,950 forks, last pushed August 17, 2026. It does what the reel implies: captures what your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions, giving Claude Code persistent cross-session memory. That's genuinely useful. What doesn't survive contact is the framing. “Unfair advantage” and “10x your next project” are vibes, not measurements — memory plugins buy you continuity, roughly the 10 minutes of re-explaining you'd otherwise do per session, and they bring their own failure modes (197 open issues; stale or wrong context getting injected is a known class of bug for memory layers). Also worth knowing: Claude Code now ships its own auto-memory, so the gap this fills is narrower than the reel implies.

A genuinely good free tool wearing a headline three sizes too big — install the tool, discard the framing.

Confidence

High

Posted by

an AI-builders page doing talking-head walkthroughs of trending GitHub repos

See the original claim →

The claim“35 CHATGPT CODES — Powerful prompts. Better answers.” — ELI5, TL;DR, SWOT, Act as [Role], Feynman, Socratic, Devil's Advocate, compare-in-a-table, JSON/Markdown output, and thirty more.

Every item on the card is a legitimate, working prompting pattern: ELI5, TL;DR, step-by-step, compare-in-a-table, pros/cons, SWOT, role prompting, the Feynman technique, Socratic questioning, devil's advocate, and output-format requests all genuinely steer any LLM — ChatGPT, Claude, or Gemini — toward more useful answers. As a free cheat sheet for beginners, it's fine. The problem is the word “codes.” There is no hidden command layer — these are plain-English phrases the model understands the same way it understands any sentence. Nothing here is documented as a command or parameter anywhere in OpenAI's or Anthropic's docs, nothing is ChatGPT-specific, and nothing is secret. The framing trains beginners to think prompting is incantation instead of clear instruction — which is precisely the mental model that keeps people bad at prompting.

A perfectly good beginner cheat sheet mislabeled as a cheat code — use the patterns, unlearn the framing.

Confidence

High

Posted by

an AI-prompts page recycling prompting cheat sheets as infographics

See the original claim →

The claim“17 Boring Business Ideas (That Rarely Fail)” with per-month revenue figures — Car Transport $141.7K/mo, Pest Control $60.4K/mo, Laundromat $27.5K/mo, Car Wash $22.7K/mo, Bookkeeping $17.9K/mo, and twelve more — tagged to Bizee business formation.

The numbers have a traceable origin, and it kills them. Laundromat $27.5K/mo, car wash $22.7K/mo, dry cleaning $27.8K/mo, bookkeeping $17.9K/mo, and storage $11.3K/mo are exactly UpFlip's published “average annual revenue” figures ($330K, $272K, $333K, $215K, $136K) divided by 12 — and UpFlip's article cites no source or methodology for any of them. What the reel presents as “what you make per month if you start one” is actually industry-average gross revenue for established businesses, skewed upward by multi-truck, multi-location operators. Revenue is not profit: a $27.5K/mo laundromat carries lease, machines, utilities, and debt service. And “rarely fail” is invented outright — no failure-rate data appears anywhere; service businesses fail at ordinary small-business rates. The kernel of truth is that boring service businesses are a genuinely sensible path, and Bizee (the rebranded Incfile) is a real formation service. Everything the numbers imply is fiction.

Traced the numbers to their source and the source doesn't support the claim — real business category, fabricated income promise, and the ad wants your formation fee.

Confidence

High

Posted by

a Bizee-sponsored reel from a page with its name conveniently hidden by the overlay

See the original claim →

The claimInfographic explaining the difference between AI, ML, DL, GenAI, LLMs, RAG, and Agentic AI — “Understand the AI Ecosystem, From Intelligence to Autonomous Action” — plus key modern concepts: multimodal AI, RAG, fine-tuning, prompt engineering, AI safety, agents.

This is one of the rare purely educational posts, and it holds up. The taxonomy is the standard, textbook-correct nesting: AI is the umbrella, ML a subset of AI, DL a subset of ML, GenAI a family of deep-learning models, LLMs a type of GenAI, RAG a retrieval technique bolted onto LLMs, and agentic AI is LLMs given tools and autonomy to act. Listing Claude among the LLM examples is accurate, and the key-concepts list matches how the industry actually talks in 2026. Two honest quibbles instead of a debunk: the “evolution flow” framing implies each stage replaced the last, when in practice they coexist — plenty of production ML is not deep learning, and most “agentic AI” is still an LLM in a loop, not a new species of intelligence. And a seven-term diagram in a 15-second reel flunks the memorability bar; nobody retains this from a scroll-past.

Verified accurate — a legitimately correct explainer chart, docked only for a format nobody learns from.

Confidence

High

Posted by

an AI-education page posting dense explainer infographics

See the original claim →

The claim“15 Top Claude Connectors You Should Start Using” — Gmail, Google Calendar, Notion, Zapier, Airtable, Stripe, Gusto, Granola, Zoominfo, Mercury, Gamma, Canva, Figma, Slack — plus a “Join AI Founder's Community” banner.

The facts check out, which puts this in the top percentile of AI reels. Claude's Connectors Directory is real — live since July 2025, holding 400+ connectors across roughly 30 categories as of August 2026 — and every single tool named in the graphic is an actual directory entry: Gmail, Google Calendar, Notion, Zapier, Airtable (read/write), Stripe, Gusto (payroll data, even running payroll for eligible customers), Granola meeting notes, ZoomInfo, Mercury, Gamma, Canva, Figma, and Slack, several confirmed on Anthropic's own connector pages. The small print the graphic skips: the headline says 15 but only 14 tools are visible in frame; “top” is arbitrary editorial, not a ranking; each connector needs per-account authorization; some capabilities (like Gusto payroll execution) are gated to eligible customers; and access depends on your Claude plan. And of course the actual payload is the “Join AI Founder's Community” banner — the list is bait, but it's accurate bait.

Verified accurate minus a miscount — a real feature list doing honest work for a community funnel.

Confidence

High

Posted by

an AI-founders community page using feature listicles as membership top-of-funnel

See the original claim →

The claim“3 Most Powerful AI Skills You Must Learn in 2026” — AI automation (Zapier, Make, ChatGPT, Notion AI, Sheets AI), AI content creation (Canva AI, InVideo, Copy.ai, Jasper), AI data analysis (Power BI AI, Tableau AI, Excel AI). “Learn. Practice. Apply. Earn... unlock unlimited opportunities.”

The nouns are fine. Every named tool exists and does roughly what the category says: Zapier and Make are real automation platforms, Notion AI and the Sheets/Excel AI features (Gemini in Sheets, Copilot in Excel) are shipping, Canva AI, InVideo, Copy.ai, and Jasper are established content tools, and Power BI Copilot and Tableau's AI features are real, if loosely named. The three lanes — automation, content, analysis — genuinely are among the most employable AI skill areas. The verbs are the problem. “Learn. Practice. Apply. Earn” compresses months of unglamorous practice into a slogan, and “unlimited opportunities” is astrology. Knowing a tool's name is not a skill; a portfolio of working automations or published content is. And the graphic quietly equates “AI skill” with “subscribing to SaaS products” — the durable skill (framing problems, checking outputs, knowing when the AI is wrong) is tool-agnostic and never mentioned.

Accurate tool taxonomy, unfalsifiable payoff — take the lanes seriously and the slogan not at all.

Confidence

High

Posted by

an anonymous listicle page running an Independence Day–themed AI graphic

See the original claim →

The claim“Claude can actually build you a business in 24 hours. Copy these 10 prompts to turn your idea into income today” — prompt pack gated behind commenting.

The true part: Claude is genuinely good at the paperwork layer of starting a business. In a day it can draft a positioning statement, an offer, landing-page copy, an outreach script, and a simple plan — and a well-written prompt does beat “help me start a business.” Everything past that is the claim, and the claim fails on what a business is. A business is not documents; it is a paying customer, and no prompt produces one. “Income today” means someone finds, trusts, and pays you within 24 hours of having an idea — that's not a Claude capability, it's a sales miracle. The 10 prompts are interchangeable with free prompts published everywhere; the artificial scarcity plus comment-gate is the tell that the product is the audience, not the prompts. The engagement math says it plainly: 241 comments against 309 likes is a gated-funnel signature — 241 harvested leads, each one now in a DM pipeline.

Killed the 24-hours-to-income claim; logged the honest version — Claude compresses the paperwork day, and the customers still take weeks.

Confidence

High

Posted by

a bathrobe-and-coffee talking-head running a prompt-pack comment-gate

See the original claim →

The claim“Top AI Tools to 10x your Income” — Lovable (websites/apps), Manus (agents), Claude (“Solves Anything”), Midjourney (art), Runway (video), Gamma (PPTs), Suno (music), Granola (notes), Cursor (code), HeyGen (avatars), Perplexity (research).

The tool list is legitimate and the category mapping is fair: Lovable really is an AI app builder, Manus a general-purpose agent, Midjourney art, Runway video, Gamma decks, Suno music, Granola meeting notes, Cursor an AI code editor, HeyGen avatars, Perplexity research. As a “know these names in 2026” flashcard, it's accurate — this is the noun done right. Two things fail. “10x your Income” is the headline and it is pure vapor: no mechanism, no example, not even a sketched path from “subscribed to Suno” to “earning 10x.” A tools list multiplies nothing. And “Claude: Solves Anything” is the over-broad billing that sets people up for disappointment — Claude is an assistant with real limits: it can't verify facts it can't check, can't do your selling, and is confidently wrong sometimes. Ironically, the honest version of this card would still be worth saving.

Accurate flashcard, fictional multiplier — learn the names, ignore the math.

Confidence

High

Posted by

an AI-tools flashcard page trading in save-for-later listicles

See the original claim →

The claim“WhatsApp Advance Agent” — a huge n8n-style node graph labeled “Group WhatsApp Agent,” presented as a running WhatsApp automation project.

The genre is fully real. n8n ships an official WhatsApp Business Cloud node, AI-agent nodes can sit in the middle to classify, summarize, and answer messages, and Meta even launched an official Groups API for the WhatsApp Business Platform. Sprawling node graphs like the one shown are what real n8n builds look like — this is a working category of project, not vaporware. The unshown fine print is where it turns. Meta's official Groups API is tiny by design: groups max out at 8 members, require an Official Business Account, and skip interactive messages. So the “group agent” reels you see running in big normal WhatsApp groups are almost always built on unofficial session-based APIs (Evolution API, Whapi, Baileys) — which sit outside WhatsApp's terms and carry a real account-ban risk, especially at volume. And a giant node graph is a flex, not a proof: the hard part — an agent that answers correctly and doesn't spam a group — is invisible in a screenshot.

Real and buildable — with the one platform detail that decides everything left out of frame.

Confidence

High

Posted by

an automation freelancer flexing an n8n node graph as proof of work

See the original claim →

The claim“My Top 7 Claude Connectors” — Notion, Slack, Fireflies, Google Drive, Gmail, Claude in Chrome, Zapier — captioned “Still paying for 5+ AI tools? You probably don't need to.”

All seven connectors exist and do roughly what the card says. Anthropic's Connectors Directory, built on MCP, ships dozens of first-party and partner connectors including Notion (read/edit pages), Slack (read channels, summarize threads), Fireflies (search meeting transcripts), Gmail and Google Drive (search mail and files, draft but not send), and Zapier (trigger automations across thousands of apps from a prompt). Claude in Chrome is a real browser extension that can read and act on the page you're on. The card even slightly undersells Gmail — accurate in both directions. The caption is the stretch: connectors let Claude read and act on your data in those apps; they don't replace the apps. You still pay Fireflies to record, Zapier still bills per task, Notion is still Notion. Most connectors assume a paid Claude plan, and each needs a per-service auth setup the reel never shows. It consolidates the chat layer, not the tool bills.

Verified accurate card with one caption-sized stretch — the connectors are real, and so are your remaining subscriptions.

Confidence

High

Posted by

an AI-training page (ai-automation.in branding) using a connectors card as course top-of-funnel

See the original claim →

The claimA copy-paste Claude mega-prompt (“The Faceless Income Plan Framework”) that outputs a 12-section plan — setup guide, 14-day action plan, “income scaling blueprint $100 to $1,000+/month” — plus 8 “Bonus Secrets” for building faceless online income.

The prompt will do exactly what a prompt can do: Claude will happily generate a structured, plausible-sounding 12-section business plan with a 14-day checklist. Structured role-plus-format prompts genuinely produce better-organized output than one-liners — that's the one checkable claim here, and Anthropic's own prompting docs back it. Faceless content businesses (niche pages, digital products, affiliate) are also real models people do run. But look at where the money lives: the “$100 to $1,000+/month scaling blueprint” exists only inside the prompt's requested output format. It's a number Claude will invent on the spot with zero knowledge of your niche, skills, audience, or available hours — not a forecast, not verifiable, not connected to anyone's actual earnings. The “Bonus Secrets” are the same recycled advice every such prompt returns. The output is a plan; the reel grades the document as if it were money arriving.

A working planning prompt with an income figure smuggled into its own output format — brainstorm with it, don't bank on it.

Confidence

High

Posted by

an AI-prompts page (RANA AI LAB branding) posting mega-prompt cards

See the original claim →

The claim“Non-engineers are shipping production apps now. Security teams are the last line of defence when something goes wrong. The CISOs at Lovable and Supabase helped us write the checklist for that.” — linking a downloadable “CISO Vibe Coding Checklist for Security.”

Everything checkable checks out. The checklist exists at aikido.dev as a real downloadable report covering technical guardrails, AI-specific controls, and org policy. The Lovable/Supabase CISO involvement is corroborated by a joint Aikido–Lovable–Supabase security masterclass webinar, and Aikido has a genuine product integration with Lovable (agent-based pentesting) — these are working partners, not name-drops. The premise is also independently real: non-engineers shipping production apps and bypassing security defaults is a documented, widely reported problem, to the point that Supabase published its own vibe-coding security checklist. The only note worth logging: it's a lead magnet — an email-gated download from a vendor that sells the solution to the problem it describes. That's standard content marketing, but the content is substantive, the named experts are accountable, and the sales pitch stays off the poster.

Verified end to end — honest vendor content about a real problem, worth the email it costs.

Confidence

High

Posted by

Aikido Security (verified page) — a security vendor whose product addresses the problem it describes

See the original claim →

The claimReel promising “better results from Claude AI without writing longer prompts” — the specific technique lives inside the video (capture frame unrecoverable).

The underlying premise is sound and vendor-documented: prompt quality beats prompt length. Anthropic's own prompt-engineering docs push structure — roles, XML tags, examples, prefilled formats — and features like Projects with custom instructions and Styles exist precisely so short prompts get better results without retyping context every time. Whatever specific tip the video carried, the genre of advice is legitimate. But the post is unverifiable as posted, and that's not an accident — it's the pattern. The actual technique is locked inside a video whose frame rendered black in our capture, which mirrors the viewer's experience: a curiosity-gap hook where the checkable content is deliberately out of reach until you watch and follow. 84 likes and 25 shares on a claim nobody can evaluate from the post itself. Graded on what's visible: a real premise wrapped in an engagement-bait delivery.

Right premise, unverifiable delivery — the honest version is one free docs page and a Project setup.

Confidence

Medium

Posted by

an AI-tips creator running curiosity-gap reels

See the original claim →

The claim“25 TOP AI TOOLS REPLACING HARD WORK — work smarter, not harder,” sorted into Ideas / Writing / Website / SEO / Image / Video categories — Behance, Dribbble, Pinterest, Drippi AI, WordStream, Madgicx, Imagica, Namecheap, Framer, Wix, GitMind, SeaArt, Stability AI, Soretext, Shuffll, and more.

Some entries are real AI tools doing roughly the claimed job — Stability AI and SeaArt for image gen, Pebblely for product shots, vidIQ for YouTube optimization, Framer's AI site builder, Shuffll and Topview for video, TextFX for writing experiments. But the list collapses on inspection. At least one tool doesn't exist as named: “Soretext” returns nothing anywhere — likely a garbled “Sora,” which isn't a caption tool anyway. Multiple entries aren't AI tools at all: Behance, Dribbble, and Pinterest are portfolio and inspiration galleries; Namecheap is a domain registrar; Wix is a site builder with an AI feature bolted on. And the categories are wrong wholesale: WordStream and Madgicx are paid-ads platforms filed under “Writing,” Drippi AI is Twitter cold-outreach filed under “Writing,” GitMind is a mind-mapping app filed under “SEO,” Imagica is an app builder filed under “Website,” Balloonary makes ad creatives, filed under “Video.” This is a scraped-and-shuffled listicle nobody test-drove — and “replacing hard work” is the exact claim none of it makes good on.

Failed the existence check — a scraped tool dump with a ghost entry, miscategorized throughout, replacing no work for anyone.

Confidence

High

Posted by

a work-smarter listicle account shuffling scraped tool names into categories

See the original claim →

The claim“My name is Kels. 7 weeks ago I had 0 followers. I trained Claude to create my content and grew to 17K.” Prompts promised in the comments (comment-gate).

The mechanism is real. Claude genuinely can batch-draft short-form content, and “training” it — really, a Project loaded with voice instructions, content pillars, and example posts — measurably improves consistency. Fast follower growth on a new faceless money-niche page is possible when reels hit, and the account behind the claim does exist on Instagram. What's on the honor system is the causal chain. “0 to 17K in 7 weeks” has no third-party verification — the claim exists only in the poster's own on-screen text, and follower counts can be bought or bootstrapped by exactly the engagement-bait this post is. Look at the ratio: 341 comments against 198 likes, mostly people typing a keyword to get “the prompts,” which inflates reach and feeds a DM funnel — the standard downstream for “digital wealth creator” bios is a course or master-resell product. And the attribution is backwards: AI wrote the words; the growth, if real, came from posting volume, hooks, trending audio, and the money niche. Claude doesn't make the algorithm push you.

Real mechanism, unverifiable trajectory, and the growth engine on display is the comment-gate itself — copy the workflow, not the math.

Confidence

High

Posted by

a faceless “digital wealth creator” persona running a prompts comment-gate

See the original claim →

The claim“Claude Code + LinkedIn Infinite Warm Outreach” — how agency owners generate warm leads from LinkedIn who already know them, on autopilot.

The kernel is real: Claude Code genuinely can research a prospect and draft sharp, personalized outreach — AI-assisted personalization is a legitimate workflow that beats template blasts. But the two words in the headline eat each other. Automated messages to people who don't know you are cold outreach with better wallpaper; “warm” describes a relationship, and no script manufactures one. Worse, the delivery mechanism is the part that gets you hurt: wiring any automation into LinkedIn — scraping profiles, bot-sending messages — is explicitly prohibited by LinkedIn's User Agreement §8.2, courts have upheld that the agreement unambiguously prohibits scraping, and accounts that scale outreach fast get CAPTCHA-checked, invitation-blocked, restricted, or banned. The one platform detail that decides whether this play survives is the one the reel never mentions. Even the market shrugged: 6 likes, 1 share.

Real drafting workflow bolted to a banned delivery mechanism — automate the research and writing, never the sending.

Confidence

High

Posted by

an agency-growth page pitching autopilot LinkedIn lead generation

See the original claim →

The claimHead of Claude Code: “I'm not prompting my agents anymore, I'm building graphs and loops so they can build the agents for me.” In 10 minutes he shows exactly where agentic engineering is going — “definitely something you cannot afford to skip.”

The person and the substance are real. Boris Cherny is Anthropic's Head of Claude Code — he created it in September 2024 — and he has said, on the record: “I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do.” His loop-engineering framing — source code → agent → loop, agents managing agents — is a genuine, documented shift in how Anthropic's own team works, and it's worth your attention. The problem is the quotation marks. Cherny's verified line is about loops prompting Claude; the tweet rewrites it as “building graphs and loops so they can build the agents for me” — splicing in “graphs,” which is the account's own coined term from its July 24 “Graph Engineering” article — and presents the blend as verbatim. The attached video funnels to the author's content, and “cannot afford to skip” is stock FOMO. Real trend, doctored quote: the account put its marketing vocabulary in an Anthropic executive's mouth.

Verified the trend and caught the splice — Cherny said loops; the “graphs” were sold separately.

Confidence

High

Posted by

an AI-content account marketing its own “graph engineering” framework off a borrowed quote

See the original claim →

The claimA company can be run by 137 AI agents mapped across sales, marketing, customer, intelligence, deals, and back office — cold email drafting is “fully autonomous” and replaces a $60–80k/year SDR or a $2–4k/month agency. Comment “MAP” for the complete breakdown.

The concept is real and genuinely useful: mapping every role in a company as an agent org chart, with each node graded on an autonomy ladder (human-led → human-assisted → fully autonomous), is exactly how serious AI adoption gets planned. Every individual agent type shown — knowledge base, cold email drafting, lead scoring, data enrichment, email verification — is standard and buildable today. What isn’t shown: any evidence a real company runs on this. The map’s own pages mark agents ‘In development,’ and the fine print under ‘fully autonomous’ cold email says a human approves every sequence before launch. It’s a diagram of 137 job titles, not 137 employees.

Built the honest version: an interactive Agent Company Map — same org-chart-of-agents concept with autonomy ladders, ‘what it replaces’ cards, and truthful Live/In-build/Planned statuses — seeded with agents that actually exist and usable as a client-facing planning and sales asset.

Confidence

Medium

Posted by

an AI-automation creator running a comment-to-DM funnel

See the original claim →

The claimA 1-person startup doesn’t need a 10-person team — it needs 7 dashboards: public status page, live changelog, metrics & analytics, waitlist counter, revenue dashboard, backend health monitor, frontend performance. Comment “10” for the full guide.

All seven dashboards are real, standard, and mostly free — none of them require the gated guide, and none require the poster’s product. What’s oversold is everything around them: dashboards show you state, they don’t run anything, so ‘like a much bigger company’ is theater. A solo founder before traction genuinely needs about three of the seven. And the comment gate isn’t generosity — it’s a DM-automation funnel run by the vendor whose product is built for making dashboards.

Built the ungated guide — all 7 dashboards, the free tool for each, real setup time, and a NOW / LATER / SKIP triage for a solo founder. No comment required.

Confidence

High

Posted by

the official page of an AI app-builder startup (vibe-coding tool, AppSumo lifetime deals)

See the original claim →

The claimI got 41.9 million views in 30 days using Claude and one content tool — no big team, no ad budget, 2 subscriptions total. Comment ‘BLITZ’ for the free guide on packaging this whole system as a service and selling it to clients.

The stack is real and the numbers are probably real — for her. Claude for content generation plus a repurposing/auto-posting tool that pushes to 8+ platforms is a legitimate, functional, cheap workflow (~$50–130/mo all-in). Her own PR cites 33M+ monthly organic views and 250+ posts per week, so 41.9M in a hot month tracks. What the reel omits is everything that actually produced the views: a ~3M-follower audience built over two years of daily posting, industrial posting volume, and the fact that she is the founder and CEO of the tool she's counting as one of her ‘two subscriptions.’ This is founder-led marketing dressed as a neutral case study. A cold account running the same two subscriptions will not get 41.9M views — it will get a content pipeline, which is still worth something.

The underlying capability — turn one long-form piece into platform-native posts everywhere, cheaply — is genuinely useful and testable in a weekend. Handed over the honest version below, minus the reseller funnel and the 41.9M expectation.

Confidence

High

Posted by

the founder/CEO of the content-automation SaaS named in the post — a ~3M-follower AI educator (199K views on this reel)

See the original claim →

The claimMost people are building agents the way that's about to be obsolete: single-agent loops break and go ‘goal blind.’ The shift is Graph Engineering — nodes, edges, state, policy, three topologies, and Anthropic's 5 official workflow patterns.

The architecture is real; the obituary is not. ‘Graph engineering’ is the mid-July 2026 X rebrand of stateful agent orchestration — LangGraph-style graphs, and structurally just state machines, a decades-old pattern. Anthropic's 5 workflow patterns are real and free to read. But ‘loops are about to be obsolete’ is flatly contradicted by the people who build this for a living: LangChain's own team moved deep research FROM predefined graphs BACK to an agentic core loop, GPT Researcher made the same move, and the best-selling agent product on earth (Claude Code) is a single-agent loop. The genuinely useful punchline — determinism via verifiers and code fallbacks beats agent count — is buried under the hype framing. Even LangChain's creator publicly shrugged at the term: ‘isn't it basically just LangGraph?’

Nothing to build — the honest version is a decision rule, handed over below. The source material worth reading is free and takes 30 minutes.

Confidence

High

Posted by

a web3/AI engagement account, ~37K followers — repackaging a viral explainer into a thread funnel

See the original claim →

The claimTurn Claude Code into a full video production studio: generate professional marketing videos through conversation — no After Effects, no Premiere, no hiring, no video editing experience.

The core workflow is real, public, and reproducible: Claude Code plus Remotion, a ‘React for video’ framework with an official Claude Code skill. You describe the video in plain English, Claude writes the components, Remotion renders an actual MP4 — setup to first video in about an hour. The creator even published his exact seven-step recipe and showed his unpolished first attempt, which is rarer than the capability itself. The only inflation is the word ‘full’: this is a motion-graphics and explainer-video studio, not a tool for editing filmed footage or generating cinematic AI video.

Handed over the exact local setup — needs Node.js and Claude Code on the user's own machine, so this is a 7-step recipe rather than a build in this session. A 2-person shop qualifies for Remotion's free commercial tier, so the incremental cost is zero.

Confidence

High

Posted by

an AI content creator (~19K views on the reel) who publishes the same workflows on a blog and YouTube channel

See the original claim →

The claimClaude Cowork is the AI breakthrough you've been waiting for (but almost everyone is using it wrong) — here are 45 workflows across productivity, writing, ops, and personal planning, from “Morning CEO Brief” to “Harsh Editor.”

Claude Cowork is real and worth knowing about: it's Anthropic's agentic mode for non-coding work — multi-step file, research, and document tasks — included with paid Claude plans since spring 2026, no separate purchase. Anthropic's own numbers say over 90% of Cowork usage is non-development work, which genuinely is a shift. The carousel's 45 “workflows,” though, are mostly well-organized prompts — “summarize my calendar into a 5-minute brief,” “critique this writing like a professional editor” — solid starters, nothing that requires the breathless framing, and slide 10's “tools for better work” list is Notion, Slack, and Google Drive, which is padding in a trench coat. The account belongs to a fractional-CPO agency; the carousel is competent top-of-funnel content marketing. Take the free prompts, skip the breakthrough framing.

Confidence

High

Posted by

a fractional-product-executive's agency account — the carousel is top-of-funnel for advisory services

See the original claim →

The claimThree repos for Claude Code: Scrapling (adaptive scraping, ~915k monthly downloads, GitHub trending #1), nanobrowser (Chrome extension giving Claude real browser control — “form fills, multi-step logins, clicking through paywalls”), career-ops (open-source AI job search, “you got the job and it didn't cost you a thing”).

The facts survive contact: Scrapling is a real adaptive scraping framework (~74k stars) whose PyPI downloads last month were 936k — the slide's 915k was accurate when made — and it verifiably hit #1 on GitHub trending in February. Nanobrowser is a real open-source Chrome extension (~13.6k stars) that works with Claude models and honestly bills itself a free alternative to OpenAI Operator. Career-ops is the same 63.9k-star repo our entry #106 verified. Two honest cautions instead of a debunk: “clicking through paywalls” is the one use-case on these slides that isn't a feature, it's a terms-of-service violation wearing a demo — the tool automates a browser, it doesn't make paid content free. And browser-agent extensions run with your logged-in sessions; point one at your accounts with the same care you'd give a new employee with your passwords.

Confidence

High

Posted by

an AI-tools creator who uses the tools for local-lead scraping — unusually, the claims check out

See the original claim →

The claimA slideshow of XML-structured Claude/GPT prompts to find your one-person business idea, build the offer, and build a sales system with “automated pipelines that attract, nurture, and convert ideal clients — without sales calls, cold outreach, or a sales team,” converting “content viewers into paying clients automatically.”

As prompts, these are genuinely decent — role/task/steps/rules structure, honest constraints like “weakest ideas flagged honestly” and “leave a bracket for real numbers, do not invent one.” Borrowing a known creator-economy philosophy for the framing is fair game. The hype lives in the outputs the prompts promise: a sales system that converts viewers into clients “automatically,” with “zero sales calls to function.” A prompt produces a plan for that system, not the system — the funnel pages, the email list, the offer that converts, and above all the audience are months of unpromptable work. For the solo businesses this targets, the no-calls promise is also strategically backwards: at low volume, one conversation closes what forty automated nurture emails won't. The prompts cost nothing and teach structure; the implied outcome is the same automation mirage this feed sells in every format.

Confidence

High

Posted by

a prompt-pack account (2,374 saves on this carousel)

See the original claim →

The claimAnthropic just released “the manual for setting up a company without employees” — watch this 30-minute talk, “don't just save it.”

Three real things, stitched into one false impression. Real thing one: Anthropic did publish a free ~35-page “Founder's Playbook” in May 2026 — but it teaches founders to orchestrate AI agents so they can hire later and smaller. It never says “no employees”; Anthropic's own Project Vend research says agents “are not there yet” on running a business alone. Real thing two: the footage is a genuine talk from Code with Claude, Anthropic's developer conference. Real thing three: the speaker is a real Google Cloud developer-relations engineer — whose 30-minute session is about deploying Claude models on Vertex AI, a technical partner talk with nothing to do with company-building. So the headline names a document the video never shows, over a talk about something else, escalated past what the document says. 10,300 people saved it; the number who noticed the speaker's job title is on his slide is presumably smaller.

Confidence

High

Posted by

an AI-news repost account (10.3K saves on this one — the save-bait worked)

See the original claim →

The claimYou're doing 1 post a week and wondering why the phone is quiet — Claude posts 20x a day for me and I write none of them. The whole system (your voice, images made, everything published for you) is Claude + Higgsfield + GoHighLevel, “about a hundred and twenty a month for all three.” Comment CONTENT for the guide and live workshop.

The arithmetic almost works and the system doesn't. Claude Pro ($20) + Higgsfield Starter (~$15) + GoHighLevel Starter ($97) is $132 on monthly billing — close enough to “about $120.” But those tiers can't do the claimed job: Higgsfield Starter's ~200 monthly credits die long before 600 posts of AI imagery, GoHighLevel bills its AI features as usage on top of the $97, and Claude doesn't publish anything by itself — the posting happens through GHL's scheduler plus glue you build and babysit. Realistic cost for the claimed volume runs $200–400+. The deeper problem is the strategy: every 2026 frequency study caps useful posting at 1–5 a day, TikTok is actively removing mass-produced AI content, and Meta demotes accounts that spam unoriginal material. And the funnel writes its own disclosure: GoHighLevel pays 40% recurring commission per referral, Higgsfield up to 15% for 12 months — one workshop conversion is worth more than this account's entire reach. The prompt in the caption, ironically, is decent.

Confidence

High

Posted by

a ~660-follower automation creator running a comment-CONTENT workshop funnel

See the original claim →

The claimTen useful AI tools (the slides show seven): Hi3D (any image into a printable 3D model), Zinley (give your AI a phone number and email), Ctruh Studio (no-code 3D/XR experiences), Terminal Candy (Mac terminal skins), gesture.live (make music with hand gestures), Wondering (Duolingo for anything), Atlaso (one shared memory layer across your AI tools).

Every tool on the slides exists and does roughly what's claimed — that alone beats most of this genre. The packaging is where it slides: the title promises ten and delivers seven; the tools presented as playthings are mostly gated (Hi3D's free tier is 100 lifetime credits with capped downloads, Atlaso's free plan covers exactly one device and one tool — the “every AI tool, automatically” magic is the $10/mo Pro plan, and it syncs your working context to their cloud), and Ctruh is enterprise contact-sales software wearing a consumer-tool costume. The link in the caption goes to an email-capture page — a free “Top 100 AI Websites” newsletter signup — because the real product of accounts like this is the list you join, not the list you read. gesture.live, credit where due, is genuinely free and genuinely delightful.

Confidence

High

Posted by

an AI-tools list account funneling to an email-capture link page

See the original claim →

The claimAn investor, customer, or co-founder gives you about five seconds before hesitation reads as doubt. Four questions cover it: who exactly is it for, what problem does it kill, why you, and what happens next. Practice all four out loud today.

The advice itself is fine — it's a compressed version of standard pitch coaching, and “if an answer takes you longer than five seconds you've found this week's work” is a genuinely useful reframe. Nothing here is measurable enough to be false. The context is what earns the flag: the account belongs to CoBuildr, a startup-idea discovery/validation tool that is itself pre-launch — no public pricing, a private-founders-program waitlist, and a site that intermittently didn't resolve when we checked. Solid borrowed advice is the content-marketing wrapper; the product underneath hasn't faced its own four questions in public yet.

Confidence

Medium

Posted by

the content account of a pre-launch startup-idea-validation SaaS

See the original claim →

The claimGive Gojiberry your ICP and it surfaces people already showing buying signals — the top 5% most active in your market, fresh job-changers, accounts in a strategic window. “No more manual list building.” Ad card: 3x more replies on LinkedIn, 4.9/5 rating, 1,000+ companies using it daily.

The product is real: GojiberryAI is a Y-Combinator-backed startup (~$99/mo Pro tier, founded 2025) that monitors LinkedIn engagement and job-change signals, scores them against your ICP, and auto-personalizes outreach — a legitimate, shipping tool in a real category. The ad's numbers are where it thins out. The 4.9/5 comes from G2 with fourteen reviews; Trustpilot — where the unhappy customers went — reads 4.3 with complaints about targeting quality and a charged-after-cancellation report. “3x more replies” traces only to the founder's own statements; no independent test exists. “1,000+ companies daily” is self-reported. And the opener is a tell: “Claude showed me a website” — Anthropic's Claude doesn't recommend products; that's a scripted credibility device, the 2026 version of “my friend the doctor swears by it.” The ad also never mentions that LinkedIn's User Agreement prohibits third-party automation, and accounts running tools like this get restricted.

Confidence

High

Posted by

the vendor's own ad account — a YC-backed LinkedIn signal-monitoring SaaS

See the original claim →

The claimSix sites for web-design inspiration, one job each: httpster.net for volume (“three thousand live sites”), curated.design for filters, hoverstat.es for the weird stuff, unsection.com for single sections (“four thousand”), pageflows.com for user flows, cosmos.so for keeping what you find.

This one holds up. All six sites are real and do what the slides say: Httpster's own homepage counts 3,116 sites (the “three thousand” claim is honest), Unsection advertises 4,000+ sections, Hoverstat.es literally brands itself “the home of alternative web-design,” and Cosmos is a legitimate visual-bookmarking tool. Two footnotes the slides skip: curated.design now redirects into the Craftwork marketplace (the standalone brand is gone, the inspiration library survives), and Page Flows is a paid product — $2.95 trial, then roughly $8–13 a month — so “open it when you need a flow” really means “subscribe first.” Also, Unsection gates deep browsing behind a free account. Accurate list, minor paywall omissions, nothing sold.

Confidence

High

Posted by

a designer/developer creator who sells web-design services — the list itself is the content

See the original claim →

The claimSix GitHub repos worth your time: Hermes Agent (229k stars, “feels 10x faster than Claude Code”), Agent Browser (40.4k), Career-Ops (63.5k), Crawl4AI (77.8k), yt-dlp (184k), Headroom (65.9k — “20% fewer tokens for coding agents”).

Rare for the repo-listicle genre: the star counts are real. We checked every one against the GitHub API — Hermes Agent genuinely sits at ~231k stars, Agent Browser at ~40.7k, Career-Ops at ~63.9k, Crawl4AI at ~78k, yt-dlp at ~184.6k, and Headroom at ~66.4k (the slide's owner slug is a typo — it's headroomlabs-ai/headroom, not labs-ai). The descriptions match what the repos actually do. The one invention is the punchline: “feels 10x faster than Claude Code.” Nobody credible measures that — the maintainer's own claim is about 40% faster on tasks once the agent has built up 20+ skills, and independent 30-task head-to-heads land at rough parity, with Claude Code ahead on output quality and Hermes ahead on autonomy. 10x is a felt number doing benchmark cosplay.

Confidence

High

Posted by

an AI-news slideshow account (vacation-photo background format, follow-for-more funnel)

See the original claim →

The claimVibe-coded apps ship with a handful of security basics missing, and that gap causes the majority of their security problems — check your app before you ship it. (“How many does your app have?” / “Did I miss anything?”)

This one holds up — it's anti-hype content pointing at a documented problem. Carnegie Mellon research found roughly 61% of AI-generated code is functionally correct but only about 10.5% is secure, and the pattern is well explained: AI coding tools optimize for the functional goal you state, and security is a non-goal unless you make it one. The documented breach list for vibe-coded apps repeats the same few failures — exposed secrets, missing authorization checks, no rate limiting, unvalidated input — none requiring a sophisticated attacker. TikTok blocked extraction of this video's exact on-screen list, so the Medium confidence is only about whether this creator's specific items are the canonical ones; the claim itself is settled. The definitive checklist is in the steps below — it's assembled from documented breaches and OpenSSF guidance, not from anyone's course.

Built the definitive version of what the video points at: the pre-ship security checklist for vibe-coded apps, below — plus the prevention layer (a rules file) that stops the holes from being written in the first place.

Confidence

Medium

Posted by

a build-in-public app developer posting vibe-coding security content — no visible funnel, which is rare

See the original claim →

The claim“I let Claude run my brand new YouTube channel for 30 days straight” — no editing, no scripting from scratch, just directing the AI. 28 days in: 69,800 watch hours and $4,066.75 in estimated revenue, on a channel that didn't exist a month earlier.

The revenue claim fails on YouTube's own calendar. A brand-new channel earns $0 from ads until it's accepted into the YouTube Partner Program — which requires hitting 1,000 subscribers plus 4,000 watch hours (or 10M Shorts views) first, and then a channel review YouTube itself says typically takes about a month. Threshold sprint plus a month-long review doesn't fit inside 28 days of existence, and views racked up before acceptance never pay retroactively. So the one number the post is built on — $4,066.75 — can't have come from the channel as described. On top of that, YouTube's monetization review explicitly rejects mass-produced, low-effort AI content — the exact channel type this system produces. The kernel of truth: AI-assisted “faceless” channels are a real format, Claude genuinely is excellent at scripting and ideation, and the poster even admits the ‘autonomous’ system has a human directing and reviewing every video. Real technique, invented outcome, and the actual product is the course.

Killed the claim; logged the honest faceless-channel path with real timelines — months to first ad dollar, not 28 days, and originality requirements the course won't mention.

Confidence

High

Posted by

a YouTube-automation course seller (“the system I teach”… “my students”)

See the original claim →

The claimTraxy catches likes and comments across all of LinkedIn, scores them against your ICP, and delivers a daily lead list with email and phone. “No more cold outreach. This is warm outreach at scale.” 7-day free trial, first 500 leads free.

The product is real and the technique underneath it is legitimate: Traxy is a shipping SaaS (G2 profile, docs, HeyReach/HubSpot/Slack integrations) that monitors LinkedIn engagement and flags ICP-matching engagers, and reaching people who are active on the platform genuinely beats cold lists. Three claims in the ad don't survive contact with the vendor's own docs. It doesn't watch “all of LinkedIn” — it monitors the specific profiles and posts you add (yours, competitors, influencers) plus an hourly agent scan. “Email and phone” come from credit-gated third-party enrichment, which never hits 100% — phone especially. And a like on a post is weak intent: the person never asked to hear from you, so the outreach is still cold contact in every legal and practical sense, with GDPR exposure on European leads. Real tool, real category, inflated coverage, and a compliance ceiling the ad never mentions.

Worth a controlled 14-day test on the free tier for our own prospecting — same protocol as prior tool tests, with hard kill criteria so a mediocre result doesn't linger.

Confidence

High

Posted by

the vendor's own ad — a LinkedIn engagement-intelligence SaaS with a 20% lifetime affiliate program

See the original claim →

The claimA Jarvis-style agent caught a tester's crash report, autonomously traced the bug, fixed it, ran tests, and shipped the fix — then placed an outbound phone call to its builder mid-drive to report the work and ask for more. "Everyone's assistants wait to be asked. Mine interrupts. No cuts, ever. This is his real speed."

Every component in this demo is real and buildable today. Crash-report-to-fix pipelines exist: an error tracker fires a webhook, a headless Claude Code session gets the stack trace, and trace-fix-test-PR is a documented working pattern. The phone call is real too, and it's the cheapest trick in the whole stack — any voice platform can place an outbound call from a webhook the moment the pipeline finishes. What earns the 'partly' is the framing. 'No cuts, this is his real speed' generalizes a best-case run: simple crashes with clean stack traces fix in minutes, and demos are always filmed on that class of bug. And 'shipped before calling' means production deploys with zero human review — the exact guardrail every sane team keeps, removed for cinematic effect. The proactive-agent idea underneath is genuinely the interesting part. The always-this-fast, never-needs-you wrapper is the hype.

Handed over the exact pipeline. One edit to the video's version: gate the ship behind a PR approval — which the phone call itself can collect by voice.

Confidence

Medium

Posted by

a build-in-public indie dev, no visible course funnel

See the original claim →

The claimAn Anthropic engineer said you're not supposed to prompt Claude — you're supposed to build a system that prompts itself. In 45 minutes 'she' shows exactly how Anthropic builds agents that remember, fix their own mistakes, and get smarter with every run. Beats any paid course on agents.

The idea is real and traceable. Boris Cherny — the creator and Head of Claude Code at Anthropic — has said versions of this on the record, most directly in a podcast interview: he doesn't prompt Claude anymore, he writes loops that prompt Claude and figure out what to do. The tooling behind it is real too: Claude Code ships loops, memory files, subagents, and dynamic workflows, and Anthropic publishes free engineering guides on agent harnesses. What's fake is the wrapper. This exact caption has circulated on X since June — same wording, but the video runtime mutates between reposts (4, 16, 24, 30, 40, and 45 minutes) and the engineer flips between 'he' and 'she' depending on who copied it. That's the signature of accounts recycling a template without watching anything. And true to form, no version of the post links the actual talk.

Handed over the real sources plus the three-piece loop recipe the caption gestures at: memory file, subagent split, stop condition.

Confidence

High

Posted by

a tech-clip aggregator account, ~18k followers, nothing for sale

See the original claim →

The claimYeah, Claude can build this. But it's never gonna design this. (Over a slick animated website — the argument being AI can code a site but doesn't have design taste or judgment.)

Refreshing for the feed: clearly framed as opinion, no comment-gate, no course, nothing sold. And the defensible core is real — AI has no client, no user research, no brand strategy, and no accountability for the outcome, so a human still decides what's *right* for a specific audience. The 2026 industry consensus agrees taste is now the differentiator precisely because generation got cheap. The overreach is the word 'never.' Current Claude models demonstrably produce coherent visual systems and propose distinct art directions — that IS design output, just design without a brief or a stakeholder. The honest version is 'AI designs *something*; a designer designs *the right thing for you*.' It's genuine commentary from someone whose living the claim defends — not a scam, but not a neutral observer either.

Treat AI as the production layer and keep a human on intent.

Effort

A brief takes an hour; it's the highest-leverage hour in the whole build

Cost

$20/mo DIY-with-taste; a few hundred to ~$2K for AI production + human direction; $3K–15K+ for full human design

Stack

A real brief + Claude for production + a human on intent

Confidence

Medium

Posted by

a working UX designer posting opinion commentary, no product

See the original claim →

The claimHow I make websites like these in 5 minutes with Opus 5 — Prompts Library + Claude. The biggest mistake is trying to get a perfect site on the first AI try; ready-made prompts let you test many directions fast. Comment 'Prompt' and I'll send you the link to the prompts for this website.

The tool is real and genuinely this good — Claude Opus 5 is a current Anthropic model that one-shots polished, animated landing pages, and the reel's actual advice (don't chase perfection on the first try; generate several directions and pick) is legitimate practice Anthropic itself documents. What's compressed out of '5 minutes': the iterations to get a take worth filming, and everything that makes a website a website — real copy, working forms, hosting, mobile QA, SEO, a client sign-off. The 'prompts library' is commodity content that's free in dozens of places, and 'comment Prompt' is a lead-capture funnel for a design agency, not a gift. Real tool, real technique, showcase-math timeline, manufactured scarcity on the freebie.

The entire 'prompts library' in one prompt, plus the honest timeline.

Effort

1–3 hours to a publishable small-business landing page, content included

Cost

Claude Pro $20/mo + free hosting vs the $1,500–5,000+ agency build behind the comment gate

Stack

Claude Opus 5 + your own copy/photos + Netlify free tier

Confidence

High

Posted by

a web-design agency running a comment-'Prompt' lead-capture funnel

See the original claim →

The claimLearning agentic AI is like a free path to a 6-figure corporate job. I can create your own system like mine — use it as a portfolio and land a crazy high-paying job.

The kernel is real and worth stating plainly: agentic-AI job postings exploded (roughly 100x growth in two years per Stanford's AI Index data, tens of thousands of postings averaging near $200K) and the AI-skills wage premium is genuinely large. Here's what guts the pitch: those salaries attach to senior engineering roles, hiring managers screen specifically for projects the candidate built and can defend, and the offer on the table — pay someone to build 'your' system — produces the one kind of portfolio that dies in the first technical interview question. The pitch contradicts itself in a single breath: a 'free path' that costs whatever the DM quote is. And handing a paid stranger the keys to an agent framework wired into your email, files, and accounts is a security exposure stacked on top of the career fantasy — the framework in question carries hundreds of security advisories and a marketplace where hundreds of malicious add-ons have been found.

The actually-free version of the free path.

Effort

Months of real laptop hours — the same months the salary is paying for

Cost

$0–20/mo to learn vs a DM invoice for a portfolio that can't survive one whiteboard question

Stack

One real problem + Claude agent tooling + your own hands + a public writeup

Confidence

High

Posted by

an agentic-AI seller pitching DM'd custom builds as career portfolios

See the original claim →

The claim5 GitHub repos / Claude Code plugins every AI builder needs: Graphify (maps your repo into a knowledge graph so Claude navigates instead of grepping), Impeccable (audits frontend for 24 design tells, kills AI-slop UI), Ponytail ('writes 80–94% less code, up to 75% cheaper, 3–6x faster'), Higgsfield MCP (image + video generation in-thread), and Obsidian as a persistent second brain.

Four of the five are real and genuinely useful — Graphify does build a repo knowledge graph so Claude stops grepping blind, Impeccable really does hunt AI-slop UI tells, Higgsfield MCP really does wire image/video generation into Claude Code, and the Obsidian-as-memory trick we run ourselves. The problem is the one with the biggest number. Ponytail's headline '80–94% less code' came from a flawed single-shot benchmark that the maintainer PUBLICLY RETRACTED — the rebuilt, realistic benchmark shows about 54% less code on average, ~20% cheaper, ~27% faster. The slideshow is quoting the dead numbers, and its '75% cheaper / 3–6x faster' don't match even the old claim. Real skill, modest real benefit, obsolete stat.

The picks worth installing, with honest expectations on each.

Effort

An hour to install the set; the Obsidian memory habit compounds over weeks

Cost

Plugins free; Higgsfield generation is $15–49+/mo in credits

Stack

Graphify + Impeccable + Obsidian memory; Ponytail with corrected expectations; Higgsfield only if you generate media

Confidence

High

Posted by

a Claude-Code creator running a follow funnel

See the original claim →

The claim5 GitHub repos every AI builder needs: n8n ('automate anything, self-hosted, no per-task fees'), awesome-claude-skills (curated drop-in skills), OmniRoute ('one API for 500+ models, 278+ providers, cheapest model per task'), Agent-Reach ('read X, Reddit and the web with zero API fees'), and last30days-skill ('mine Reddit for real opinions in ~20 seconds').

Better hit rate than most of these lists — all five repos exist and do roughly what's claimed. The sins are the omitted asterisks. n8n is free to self-host with no per-task fees, but it's fair-code, not open source: hosting it for multiple clients and charging crosses into needing a commercial license. OmniRoute is real and genuinely routes across hundreds of providers, but 'free' means the router is free — the models behind it still bill once you exhaust rate-limited free tiers, and piping code through 290 third-party endpoints is a data-governance decision, not a lifehack. Agent-Reach's 'zero API fees' is literally true and exactly why it's fragile: it leans on scraping and cookie-auth that violate X's and Reddit's terms and break when those sites change. Great for personal research; risky for client work.

Use all five with the caveats the slideshow left out.

Effort

An afternoon to try all five; longer to self-host n8n properly

Cost

$0 to start + n8n license if you productize + provider bills once OmniRoute's free tiers run out

Stack

n8n (your own instance) + awesome-claude-skills + OmniRoute with paid keys for anything real

Confidence

High

Posted by

a builder-content page running a follow funnel

See the original claim →

The claimSeven underrated design tools — Ditther, Cooldock, Selected, Logo System, Macfolio, Ultramock, and Javii. All linked, all underrated. Save this for later.

We hunted down all seven: every one exists, names spelled as given (yes, Ditther has two Ts), and the picks are decent — Ditther and Ultramock are genuinely fun, Logo System is a solid inspiration library at 1,200+ logos. An honest roundup by a real designer with no comment-gate and no product push. Two asterisks keep it honest: the listicle vibe implies free candy, but Cooldock is a paid app and Ditther/Ultramock gate their best exports behind Pro tiers — using all seven seriously runs $20–50, not $0. And two of the seven aren't tools at all — Selected and Macfolio are discovery directories. Padding, not deception.

The picks worth your bookmark bar, plus the free routes around the paid ones.

Effort

Ten minutes of clicking to know which two you'll keep

Cost

$0 to browse and try; $20–50 to own the paid tiers seriously

Stack

Ditther free + Logo System + shots.so covers most of what the seven promise

Confidence

High

Posted by

a product designer sharing tools, audience-building, nothing gated

See the original claim →

The claim5 open-source AI trading agent repos across crypto, stocks, and prediction markets — 'the real ones where an LLM actually reasons.' The caption itself concedes the biggest one can't place trades by design, and says to run everything in paper-trading mode first.

By TikTok standards this caption is honest — it tells you the flagship can't trade and to paper-trade first, which is more candor than the genre ever shows. The repos are real: TradingAgents (~98K stars) and ai-hedge-fund (~63K) are the two giants, and ai-hedge-fund's README says in bold that it's educational and doesn't place trades. But the receipts get worse the closer you look: TradingAgents' own community caught its backtests leaking future data (a 'clear look-ahead bias' issue — real-time info retrieved during historical simulation), one listed repo is a general agent framework rather than a trading agent, one is Microsoft research tooling, and one is an 87-star repo idle since April. No verified live profits exist for any of the five. And the flag that matters most: the poster's own site sells an AI crypto-perps trading terminal — a trading-repos post from that seller is top-of-funnel, and the caption doesn't say so.

The education is real and free — here's how to take it without paying tuition to the market.

Effort

A weekend to run paper mode; months of it before the results mean anything

Cost

$5–20/day in API calls to run seriously — the simulation bills you even when the market doesn't

Stack

ai-hedge-fund (paper) + a cheap model + a calendar reminder to still be paper-trading in 90 days

Confidence

High

Posted by

an AI-content creator (a real engineer) whose own product line includes an AI crypto-trading terminal — not disclosed in the post

See the original claim →

The claim5 open-source repos that find you paying clients, starting with local businesses that don't even have a website: a Google Maps scraper, an email-existence checker, a no-code scraper builder, self-hosted forms, and a self-hosted helpdesk — replacing what people pay Apollo, Instantly, and ZeroBounce for.

All five repos are real and maintained, and 'local businesses without websites' is genuinely a smart niche. Then the pipeline falls apart at every joint. Scraping Google Maps breaches Google's terms — the answer can be IP blocks or your Google account. The email-verification step — the hinge of the whole pipeline — is documented by its own maintainer as impractical from home: it needs outbound port 25, which residential ISPs block, and real volume needs SMTP proxy servers. Nothing in the stack does deliverability, which is the thing Apollo and Instantly actually charge for. And the promised outcome — paying clients — is the step no repo performs. The tools produce a list. The list is not a client.

The version that gets local clients without the ToS breach — it's mostly not software.

Effort

An afternoon builds a better list by hand than the pipeline builds in a week of setup

Cost

$8/1,000 verifications + a domain vs $50+/mo in proxies and the legal tail risk

Stack

Google Maps + a spreadsheet + a phone — the stack nobody posts because it isn't shiny

Confidence

High

Posted by

an AI-content creator (a real engineer) who sells 1:1 consults and app builds

See the original claim →

The claimTen open-source repos, 446,694 stars between them, doing work you'd expect to pay for: Stirling-PDF→Acrobat, Immich→Google Photos, Upscayl→Topaz, Whisper→Otter, LocalSend→AirDrop, plus five newer agent-tooling picks. Most run entirely on your own machine; costs nothing but the time to install.

We added the stars ourselves: the claimed total was accurate at posting (455K by the time we checked — they grew). All ten repos exist, and the headline five are exactly the tools the self-hosting community already swears by: Stirling-PDF (~89K stars), Immich (~110K), Upscayl (~48K), Whisper (~107K), LocalSend (~88K). These are the real deal. The honest footnotes: 'Whisper replaces Otter' compares a model to a product — Otter's meeting bot, speaker labels, and summaries need a wrapper app around Whisper; Immich makes you the sysadmin of your family's photo library (always-on box, backups, docker); and the five agent-tooling picks quietly sit on top of a paid Claude plan. 'Free' means you pay in setup time and a machine that stays on. As repo lists go, this is the good stuff.

The install order that pays off immediately.

Effort

Two are instant; the self-hosted ones are a weekend each

Cost

$0 software + a $100–300 always-on box for Immich + $2/mo to Google remains the honest lazy option

Stack

LocalSend + Upscayl + Stirling-PDF first; Immich for the committed

Confidence

High

Posted by

an AI-repos page that counts stars honestly (same account as #089)

See the original claim →

The claimUnreal Engine x Claude Code x MCP — 'this thing is built from scratch,' 'Claude Code after 2h using Unreal Engine,' shown as a photoreal GTA-style open-world city.

The footage isn't Claude's work — it's 'The Matrix Awakens,' Epic's famous Unreal Engine 5 tech demo, and the post's own audience says so in force. The top comment ('fake, it's the Matrix demo on UE') has hundreds of likes, and reply after reply identifies the same map, the same character, even the same on-screen 'fly' controls from the demo. What actually happened is a real, months-of-work Epic showcase got relabeled as 'Claude Code, from scratch, in 2 hours.' The real thing underneath — Claude driving Unreal via MCP — genuinely exists and can do useful editor work (we covered it honestly in our Geopogo entry). What it cannot do is generate a photoreal open-world city from scratch in an afternoon, and this clip is proof by misattribution, not by capability.

What Claude + Unreal actually does today — the real version is still worth knowing.

Effort

30 seconds to identify the source clip

Cost

$0 — this is a spot-the-fake, not a buy

Stack

UE 5.8 native MCP + Claude for the real (modest) version; a reverse-search habit for the fakes

Confidence

High

Posted by

a creator whose other posts promote AI web-design services

See the original claim →

The claimA story of building side income with Claude, ending in a plug for Tase — an AI life assistant that 'turns one spoken sentence into logged actions across money, health, tasks, and goals.' Screenshot shows '1.8K ratings, 5.0 stars.' Origin story: built a 5-tab tracking sheet with Claude in 8 minutes, got 7 paid requests at $15 in week one. Comment TASE for a 14-day free promo code.

Tase is a real app (Tase LLC, on the App Store) and it does roughly what's claimed — voice-to-logged-actions across money, health, and tasks. But the social proof is the problem. The slide shows '1.8K ratings, 5.0 stars'; the live App Store page shows 4.8 stars from 53 ratings — the screenshot inflates the rating count roughly 34x and shows a perfect 5.0 that doesn't exist on the real listing. The origin story (a 5-tab sheet built in 8 minutes, 7 paid requests at $15 in week one) has zero independent trace and carries a price anchor baked in — the shape of funnel fiction, not a case study. And 'free' is a weekly-subscription app: after the promo it's $3.99–6.99/week or $80–200/year.

How to test the real app without getting anchored by the fake numbers.

Effort

Minutes to try the free tier; seconds to check the real rating

Cost

'Free' app = $80–200/yr after promo vs $0 for a Claude + Sheets logger

Stack

Tase free tier to test — or Claude + a Sheet if you'd rather not subscribe

Confidence

High

Posted by

a creator funneling to a comment-'TASE' promo code for the app being advertised

See the original claim →

The claimTen-repo list led by: AutoHedge (3.8K stars) — run an AI hedge fund from your laptop, 4 agents for research, validation, risk, and trade. Plus Vibe-Trading (describe a strategy in plain English, it trades), a Bloomberg-without-the-$25K-seat terminal, self-hosted every-model chat, a free AI video studio, and five more.

We checked all ten: every repo exists, every quoted star count was accurate at posting (they've since grown), and several are genuinely excellent — LibreChat, HeyGen's hyperframes, and Vibe-Trading (a real university lab project with paper-trading mode and audit logging). The hazard is the headline verb. AutoHedge is a 37-commit marketing demo for an agent framework — its own maker calls it a demo app. Live execution is Solana-only, setup involves pasting your wallet's private key into an env file, and the README carries no risk disclaimer and zero performance results. Here's the market's own receipt: in the Alpha Arena live benchmark, six frontier models traded $10K each of real money — four of six lost, one down 63%, losses driven by the models over-trading away their capital in fees. That's the best AI in the world with a professional harness. Your laptop is not positioned to beat it.

How to get everything real this list offers without funding a wallet.

Effort

Paper-trading setup: a weekend. Understanding why the bot's confident thesis is wrong: the actual skill

Cost

$0 repos + $50–500/mo tokens if run seriously + your stake if you ignore the paper-trade step

Stack

Vibe-Trading (paper mode) + LibreChat + an index fund for the money part

Confidence

High

Posted by

an AI-repos page that counts stars honestly and oversells verbs

See the original claim →

The claimSeven Claude Code repos to save for your next build, led by ruflo — 'the successor to claude-flow, a swarm of Claude agents that coordinate and share memory across a whole project' — plus 100+ ready-made subagents, GitHub-Issues project management, PRD generators, Anthropic's official agent SDK, slash commands, and a CLAUDE.md starter. Free starter kit in bio.

All seven repos exist and five are actively maintained — the subagents collection (~24K stars), Anthropic's official SDK, and the PRD workflow are accurately described. The problem is the headliner. 'Successor to claude-flow' inflates a rename into a new product: same repo, same author, renamed at v3.5 for trademark reasons. And ruflo's own community reports the harder problem — a discussion thread on the repo describes the verification system as broken 'at a critical level,' with agents self-reporting success when most actually fail. A 67K-star swarm that grades its own homework and gives itself an A is exactly the thing you don't want shipping your project. The bio's 'free starter kit' is a standard email-capture lead magnet.

What we actually run instead of a hundred-agent swarm.

Effort

An afternoon to a solid Claude Code setup without any meta-harness

Cost

Repos $0; swarm-scale token burn is $100+/mo — simple setups run fine on Pro

Stack

Claude Code + CLAUDE.md + 2–3 verified subagents + you on the merge button

Confidence

High

Posted by

an AI-content creator running a link-in-bio starter-kit funnel

See the original claim →

The claimThis week's trending GitHub repos: an AI that applies to jobs for you, one that runs Word/Excel/PowerPoint, one that lets Claude control your entire computer, one that kills AI slop, and a free gateway to 50+ AI models. All free and open source.

All five categories are real and open source, and one pick (the Office automation tool, OfficeCLI) genuinely was trending the exact week of the post — we found it in two independent trend reports from that window. But the lead item, the job-application bot (AIHawk), had its viral moment in 2024 — recycling old viral repos as 'this week' is the genre's standard move — and auto-applying violates LinkedIn's terms, an account-ban risk the post never mentions. 'Free gateway to 50+ models' conflates free software with free model usage: the gateway costs nothing, most models behind it still bill per token. And 'lets Claude control your entire computer' is a real capability that deserves a security warning, not just hype.

The versions of these five jobs that don't cost you an account or a surprise bill.

Effort

Minutes to check any trending claim yourself

Cost

$0 for the repos; tokens, ban risk, and security exposure are the hidden line items

Stack

GitHub trending (raw source) + Claude for the resume work + gated agents over raw ones

Confidence

High

Posted by

an AI-tools creator farming saves

See the original claim →

The claimTotal monthly cost to run a startup: ~$20. Claude for coding ($20/mo), Supabase (free), Vercel (free), Namecheap domain ($12/yr), Stripe (2.9%/transaction), GitHub (free), Resend, Clerk, Cloudflare, PostHog, Sentry, Upstash, Pinecone — all free tiers. There's never been a cheaper time to build.

The individual claims are basically accurate — every tool listed is real and every 'free' really is a genuine free tier that thousands of builders launch on. As a snapshot of how cheap it is to START, the post is honest and even useful. Where it slips is the headline '~$20 to run a startup.' Free tiers are onboarding ramps, not the running state: Supabase, Vercel, Clerk, Pinecone, PostHog, and the rest all have usage ceilings, and the moment you have real users you cross them and the bill climbs — often to hundreds a month. Stripe's 2.9% is also a real cost the '$20' total quietly omits. So: ~$20 to build and launch a prototype, yes. ~$20 to RUN a startup with actual traffic, no.

Use the stack honestly — cheap to start, with eyes open about what scales.

Effort

A weekend to a live prototype on the free stack

Cost

~$20/mo to build and launch; hundreds/mo once you have real users

Stack

Claude + Supabase + Vercel + Stripe + a domain — real, and priced for month one

Confidence

High

Posted by

a tech creator sharing a startup stack, no product

See the original claim →

The claimOpen-source tools are getting seriously good — in many cases they do the same jobs people pay monthly subscriptions for, across AI, automation, coding, scraping, databases, and productivity. This list of 10 GitHub projects proves it. (None of the 10 are named in the caption.)

The premise is verifiably true as a category — this audit has confirmed a dozen instances of it (Stirling-PDF vs Acrobat, n8n vs Zapier, PostHog vs Mixpanel, Baserow vs Airtable, Chatwoot vs Intercom). But 'these 10 projects' can't be receipt-checked because the caption names zero of them. Un-nameable tools are un-gradeable tools, and the vagueness is the retention mechanic: you must swipe, save, and follow. Compare entry #070, which at least named its list and could be held to it. The control/customization/lock-in argument is directionally fair and omits the same cost every time: self-hosting means you become the sysadmin.

A vetted, NAMED list in the same categories — every one inspectable before you install it.

Effort

One tool per weekend is the sustainable pace

Cost

$0 software + $5–20/mo hosting + your new sysadmin hobby

Stack

The named six above — same categories, actual receipts

Confidence

Medium

Posted by

an AI-tools page farming follows with mystery lists

See the original claim →

The claim5 open-source AI video generators most people are sleeping on — the same kind of tech behind paid tools like Runway, Kling, and Higgsfield, except free and running on your own machine.

The category is real and 2026 was genuinely its year: Wan 2.2 (Apache-licensed, Alibaba), HunyuanVideo 1.5 (Tencent), and LTX-Video produce output that independent comparisons put near mid-tier paid platforms. But this caption names none of its five repos — the list lives only in the slides, so the claim can't be pinned down as posted, and that alone caps the grade. 'Same kind of tech' is defensible (same architecture family); 'the same tech' would be false — Runway's and Kling's actual models are proprietary. And 'on your own machine' silently means a 16–24GB-VRAM GPU: a $1,600+ graphics card and minutes-per-clip render times. The paid tools' real product is speed and not owning the hardware.

The named version of the claim, plus the no-GPU route.

Effort

A day to a working ComfyUI pipeline if you're technical

Cost

$0 models + a $1,600+ GPU you may not own, or ~$0.50/hr rented

Stack

ComfyUI + Wan 2.2 (local) or a cloud GPU by the hour

Confidence

Medium

Posted by

an AI-content creator (a real engineer) who sells 1:1 consults and app builds

See the original claim →

The claimGeopogo MCP connects Claude directly to Unreal Engine: create Blueprints, modify levels, place actors, generate gameplay systems in natural language. Available on the Fab store.

Geopogo is a real San Francisco design-AI company (they also ship a Claude-to-Revit plugin), the product page and Fab-store listing exist, and the capability class is genuine — Claude driving the UE5 editor over MCP works and demos well. Two things temper it. First, 'generate gameplay systems' stretches what these bridges reliably do: the small verbs (spawn actors, set properties, scaffold Blueprints) work; a functioning gameplay SYSTEM still needs a developer steering and debugging. Second — the part the ad understandably skips — Epic itself now documents native Unreal MCP support in UE 5.8, and free open-source bridges predate this product. We couldn't verify the Fab price (the store blocks automated fetches); the company's other tools are freemium-to-paid.

How to get the demo's magic without buying first.

Effort

An evening to a working free bridge if you know Unreal basics

Cost

UE free + Claude ~$20/mo; Geopogo's price unverified — check before buying

Stack

UE 5.8 + native Unreal MCP or a free GitHub bridge + Claude

Confidence

Medium

Posted by

Geopogo — the tool company's own account, advertising its own product

See the original claim →

The claimMCPs are how you turn Claude from a chatbot into an agent that controls your browser, your repos, your files, and even Blender. Named: the official MCP servers repo, Playwright MCP (Microsoft), GitHub MCP (GitHub official), Blender MCP, Context7.

All five verified real, active, and massively adopted — and three of the five are published by Microsoft, GitHub, and the MCP project itself, which is about as official as receipts get. The core explanation of what MCP is (a protocol giving Claude tool access to external systems) is accurate. The caption's visible keyword-stuffing is an algorithm play, but the links are direct GitHub URLs with no gate and no product. What it skips is only friction: each server needs a config-file edit, a runtime, and 30–60 minutes of setup the carousel doesn't show. Second 'real' in this batch for an MCP listicle (see #072) — the MCP genre is unusually honest, probably because the tools are impressive enough without lying.

Same verdict as #072 and the same receipt: our own daily operation runs on exactly this pattern — the browser automation that fetched this post is an MCP.

Effort

30–60 minutes for your first server

Cost

Servers $0 + Claude Pro ~$20/mo

Stack

Pick ONE: Playwright MCP (browser) or GitHub MCP (repos) — add more as trust grows

Confidence

High

Posted by

an AI-prompts account with direct GitHub links and no gate

See the original claim →

The claim5 prompts stolen from a friend who's quietly outpacing everyone — Brain Dump, Voice Clone, Hard Reply, Lazy Genius workout, Regret Test. Most people don't realize ChatGPT can do this. These aren't basic.

Credit first: the full prompt text is in the caption itself — no 'comment PROMPTS,' no gate. The prompts run fine and two of them (Brain Dump, Hard Reply) are genuinely handy everyday templates. Now the grading: these are standard role-prompting patterns that have circulated in every prompt roundup for two-plus years. 'They're not basic' — they're basic. The 'friend quietly outpacing me' is a stock engagement frame, not a source. And the expert-replacement angle is inflated where it matters: the Regret Test is not a financial advisor and the workout prompt doesn't check your form. Useful starters, oversold origin story.

Skip the mystique — the entire technique in five lines.

Effort

10 minutes to learn the pattern once, then never save a prompt pack again

Cost

$0

Stack

Free-tier Claude or ChatGPT + your own context

Confidence

High

Posted by

a prompt-tips account

See the original claim →

The claimA prompt 'jailbreaks Claude [all versions]' into writing code it would normally refuse — shown building a Discord mass-DM/'nuke' bot and a Roblox cheat script.

We grade tactics, not people, and this tactic fails on both honesty and usefulness. We did not capture, test, or reproduce any jailbreak prompt — that's a hard line — so this entry grades the CLAIM, not a recipe. The claim is that one prompt reliably unlocks 'all versions' of Claude to write malware. Its own comment section is the debunk: the top replies point out it only ran on an older model, that newer and Opus-class models don't play along, and that the people who try it get their accounts and game accounts banned. So the pitch — a universal, consequence-free jailbreak — is false on every axis: not universal, not reliable, and definitely not consequence-free. What's real underneath is only the boring, ugly part: some models can sometimes be manipulated, and the AI labs keep closing exactly these holes.