The AI Hype Audit

What holds up

• Granola genuinely captures meetings without a bot — system-audio transcription, no participant-list entry (its own docs)
• The official Granola MCP server is real (Feb 2026) and works from Claude, Claude Code, and ChatGPT — the demoed workflow is the product working as designed
• It's a properly DISCLOSED paid partnership with a link card — more than most AI ads manage
• The company is legit: $125M Series C at a $1.5B valuation, March 2026

What doesn't

• Full MCP access is a paid feature: Business plan $14/user/mo; free tier = 30 days of personal notes, no transcripts
• Bot-less capture means call participants get NO indicator they're being transcribed — an all-party-consent problem in states like CA, FL, WA, IL, PA
• Audio is processed by a cloud transcription provider — it is not local-only, whatever 'notepad' implies
• 'Cheat code' is influencer-speak for 'documented official feature'

How to actually do it

• Say 'I'm running an AI notepad on this call' at the top of every meeting. Five seconds, and it converts the legal risk to a professionalism flex
• Budget the $14/user/mo Business seat if you want the full Claude integration — the free tier's 30-day window will feel broken the first time you need last quarter's call
• Setup is one line: add the 'Granola' connector in Claude settings, or claude mcp add granola --transport http https://mcp.granola.ai/mcp
• Test the recall before you trust it: ask Claude for yesterday's action items and check them against your own memory of the call once

This one you can actually copy — just do it with consent and a budget line.

What holds up

• The White House AI Action Plan is real — July 23, 2025, three pillars, 90+ federal actions
• 36 million small businesses is the current SBA figure (36.2M as of June 2025) — he's more current than most skeptics
• The gap is genuinely real: Census says under 9% of small firms actually use AI in operations, and most that do have no formal strategy
• AI-implementation-as-a-service is a legitimate business — it's ours

What doesn't

• 'Directs agencies to drive AI into small businesses' = spin on ONE open-source convening provision; the plan's own text says the adoption worry is big orgs, not Main Street
• '92% no AI strategy' and '3% of consultants' — no source exists for either; the precision is costume
• The OnlineJobs.ph link card pays him a 40% recurring commission whether or not you ever buy his program
• His program has no public price — application, then a 1-on-1 'strategy session' close, aimed at '$150k+ corporate professionals'
• Same person previously sold a 'Modern SDR' course and a 'remote triaging' opportunity — new wave, same funnel
• The delivery model hands a US small business's customer data to a $6/hr offshore contractor with zero local accountability — that's where the widely-cited '95% of AI pilots fail' stories come from

How to actually do it

• Before any 'the government just changed everything' post: pull the actual document. The plan is a free 28-page PDF on whitehouse.gov — the small-business section takes 30 seconds to not find
• When a stat sounds load-bearing ('92%', '3%'), search it with the word 'survey'. If the only source is the post itself, it's decor
• Check what the poster earns from the links: OnlineJobs.ph pays 40% recurring to affiliates — an embedded link card is a revenue channel, not a resource
• The legit version: scope the business's actual bottleneck first, install and test personally, keep the data inside the business, document everything, train the owner, and price on delivered hours saved — not on a stat nobody can source

The honest version of this business exists. It's boring in the best way, and it can't be run from a beach.

What holds up

• All 6 repos verified directly on GitHub — real, public, active
• Star counts on slides accurate; most have grown since posting
• All permissively licensed (MIT / Apache 2.0)
• Poster discloses setup effort — no get-rich framing
• Repo links given away free, no email gate, no visible course funnel
• HyperFrames is HeyGen's own open-source project — corporate-backed, not abandonware risk

What doesn't

• "Replace ElevenLabs" oversold — local cloning needs a real GPU or Apple Silicon and trails on quality
• career-ops requires an AI coding CLI subscription to run — replacing one subscription with another
• World Monitor is not Bloomberg — no market data terminal, no trading; the $32k comparison is theater
• "Free" excludes compute, API tokens, and your setup hours
• Six installs is a hobby; most people need one or two of these

How to actually do it

• Agent-Reach (research agent superpowers): needs Node 18+. Run `npx agent-reach install` in a terminal, or add it as an MCP server to Claude Code — repo README at github.com/Panniantong/Agent-Reach has the one-line MCP config. Zero API fees; it scrapes directly, so respect platform ToS for commercial use.
• Hallmark (anti-slop design checks): in Claude Code run `/install github.com/Nutlope/hallmark` or copy the skill folder into ~/.claude/skills. It then auto-runs 57 design checks on any site your agent builds. Free, no accounts.
• HyperFrames (HTML-to-video): `npm install hyperframes` in a project, write scenes as HTML/CSS, render to MP4 locally. Docs at github.com/heygen-com/hyperframes. Needs Node + ffmpeg installed.
• Voicebox (voice cloning/dictation): desktop app download from voicebox.sh — needs Apple Silicon Mac or a Windows machine with an NVIDIA GPU for usable speed. Voice cloning laws apply: clone only voices you have rights to.
• career-ops and World Monitor: personal-use tools. career-ops installs into Claude Code (`/install github.com/santifer/career-ops`); World Monitor is `git clone` + `npm run dev`, or just use the hosted version at worldmonitor.app free.

Verified all six against GitHub primary sources. Flagged the three with real business value for a consulting shop — Agent-Reach (agent research without per-call API fees), Hallmark (design-quality checks on AI-built client sites), HyperFrames (agent-generated video for content) — with exact install steps. The other three are personal-use tools.

What holds up

• Real registered company (Explee LTD, London) with public pricing and API docs
• AI Search over the company database works and has a free tier — no signup required
• Pay-as-you-go pricing is real: ~$0.03/email, $50 free credits for new accounts
• Pre-warmed mailboxes and automated follow-ups are a real, shipping feature
• Email verification via ZeroBounce; sub-3% bounce claim is plausible for verified data

What doesn't

• Their own math: 1,000 emails → 1–2 meetings. A ~0.1% meeting rate is the honest number under the hype
• 97% deliverability is the vendor's own figure — no third-party audit
• All glowing testimonials live on the vendor's homepage; only a handful of G2 reviews exist
• Cold email at volume has a legal ceiling (CAN-SPAM, GDPR for EU targets) — and their terms of use push all compliance liability onto you
• Sending from vendor-managed mailboxes means your brand reputation rides on their infrastructure

How to actually do it

• Free win first: go to explee.com/landing-search (no signup needed) and run a plain-English ICP search — e.g. 'multi-truck home service companies in Northern Colorado with 10+ employees'. Judge data quality on names you already know.
• If the data holds up, create a free account: 500 welcome credits, no credit card.
• For the AutoGTM test: sign up at explee.com, use the $50 new-account credit or put in $30 — that's ~1,000 emails.
• Point it at ONE specific offer with a clear call to action, not your whole business. US-only targeting to stay out of GDPR territory.
• Set the daily budget cap low (it's in the campaign settings) and read the first batch of AI-written emails before letting it run — your name is on them.
• Success bar from their own numbers: 1–2 booked meetings per $30. If a test booking costs you under $15, it beats most paid channels; if it returns zero after 1,000 sends, kill it — the data was fine, the offer wasn't.

Worth a controlled test, not a leap of faith. The free AI Search tier is the no-risk win — use it for list-building today. AutoGTM itself gets a capped $30–50 experiment on one tight offer before any real budget.

What holds up

• Supercool is a real product that does generate music, video, voice, images, and text
• The itemized subscription prices in the post are roughly real market prices
• The company answers Trustpilot complaints and has issued refunds when pushed
• The frustration it targets is real: juggling five AI tools and being the glue is a genuine pain

What doesn't

• The 'satisfied switcher' is the product's own CEO — an undisclosed ad dressed as a testimonial
• Promoter under active SEC fraud charges (Sept 2025, ~$112M alleged scheme with Tai Lopez)
• $8 buys starter credits; auto top-up defaults ON — documented user outcomes: $534 in 20 minutes, $600+ surprise charges
• Same low-ticket-hook playbook as Mentorbox, his earlier venture with years of BBB billing complaints
• 'No AI tool beats Supercool' — it's a wrapper around unnamed third-party models; the claim is unfalsifiable
• ~239 active Meta ads, ~26 new creatives a week — you didn't stumble on a post, you were targeted by a funnel

How to actually do it

• Before believing any 'I switched' post, search '<person> + <product> + founder'. Took us one search to find the CEO title
• Never leave auto top-up on with any credit-based AI tool — that's where the $534-in-20-minutes stories come from
• If you actually want to consolidate: pick the two tools you use weekly, cancel the rest. That's the honest version of cutting $211/month
• A tool that won't name the models it runs on is telling you the markup IS the product

If a testimonial moves you, the first check is always the same: does the person talking own the thing they're praising?

What holds up

• Real, identifiable operator; real SEC-registered company with genuine Form C filings
• The crowdfunding raise, summit, and 3-day clone challenge all verifiably exist
• The Reg CF disclosure documents are honest — they even disclose his FTC order in the officer bios
• The underlying idea (AI clone of yourself for content) is real, buildable technology

What doesn't

• 2023 FTC order: $14M judgment, $2.5M refunded to ~1,900 customers, permanent ban on unsubstantiated earnings claims
• '$4M raised' is a self-issued press release; SEC filings verify ~$465K of crowd money
• '$3 BILLION projected' appears nowhere in any SEC filing — pure marketing number
• The 100-company roll-up currently consists of a newsletter and his own former company
• $0.10 illiquid Reg CF shares from a company running a $2.9M annual loss
• Long complaint trail on his programs: refused refunds, aggressive upsells — the FTC-ordered refunds confirm it wasn't anecdotal

How to actually do it

• Want the AI clone? Build it: ElevenLabs voice clone (~$22/mo) + HeyGen avatar (~$29/mo) + a Claude project loaded with your own transcripts as the brain. ~$50-80/month total, one weekend of setup — no challenge, no upsell ladder
• Before ANY crowdfunding 'want in': read the actual Form C on SEC EDGAR, not the press release. Five minutes, free, and in this case it contradicts the pitch on the numbers
• When a promoter has a federal order about earnings claims, weight every number in the ad at zero until the filing confirms it
• Rule of thumb: real raised money has an SEC accession number; 'projected' money has a countdown timer

The clone tech itself is real and cheap to build yourself — that's the honest alternative, and we priced it out.

What holds up

• The Skool Games are a real competition with real, verifiable mechanics
• Real operator with a substantial, genuine Skool footprint — not an anonymous ad account
• The free community and its templates genuinely exist and cost nothing to join
• Top-10 placement is genuinely plausible for a community his size — the credential isn't absurd, just unverified

What doesn't

• 'Winner' is doing heavy lifting: Skool calls all ten top finishers per round 'winners'
• The 2x-winner credential is self-asserted only — absent from every public winners list we could find
• Brand-new FB page, single-digit engagement; the checkmark is paid Meta Verified, not notability
• 'FREE. For now.' is manufactured urgency on a lead magnet that was always going to be free — its job is booking calls
• 'Guaranteed' scaling language in the bio; no published pricing anywhere in the funnel

How to actually do it

• Before buying any Skool Games credential: ask WHICH round and WHAT placement, then ask for the announcement link. Top-10 finishes get announced — a real winner can produce the receipt in one message
• Take the free community if the templates are useful — just know you're entering a book-a-call funnel, and the 'for now' deadline is theater
• Never book the sales call before you know the price range. 'We'll discuss on the call' means the price flexes to your pain
• The actual playbook these funnels run — free value, community, call-gate — is itself the client-getting system being sold. You're watching the lesson for free

Not a scam — a real operator with an inflatable credential. Grade the systems, not the trophy.

What holds up

• Gemini Gems are free on every Gemini plan, including the free tier
• Genuinely no-code: name, instructions, optional knowledge files, save
• Under 5 minutes is realistic for a basic Gem
• Real business value for repeat tasks — a saved specialist beats re-pasting prompts

What doesn't

• "Just made free" — Gems went free to all users in early 2025; the news is over a year old
• $5,000 agency comparison is apples-to-oranges: Gems aren't customer-facing website chatbots
• No lead capture, no site embed, no CRM hooks — the things the $5k actually buys
• "Read comments" = comment-farming funnel, the payoff is engagement for the poster

How to actually do it

• Go to gemini.google.com signed into any free Google account
• Click your profile avatar → Gems (Gem manager) → New Gem
• Name it, then write Instructions like a job description: who it is, what it does, output format. Use the rewrite wand if stuck
• Add Knowledge files (pricing sheet, FAQ, style guide) so it answers from your material
• Pick a Default tool if relevant, Save — it now sits in your sidebar as a reusable specialist
• Need the actual $5k product — a bot on your website answering customers and capturing leads? That's a different build: a chat widget wired to an AI backend with your knowledge base. Real project, not five minutes

Exact 5-minute Gem build steps below, plus what the $5k agency build actually adds so you know which one you need.

What holds up

• Real, named operator with a consistent public footprint — not an anonymous ad account
• The underlying product genuinely exists and works: a Claude-based 'brain' with named assistants and task skills is a real, buildable thing for non-coders
• The webinar is genuinely scheduled and hosted live (on a fortnightly cycle)
• Refund condition on the install tier ('working system or refund') is stated on the page

What doesn't

• 'Before spots are full' is a manufactured-scarcity counter on an evergreen fortnightly webinar
• Zero independent evidence anywhere: no testimonials, no reviews, no student results
• 'Runs your day' autonomy takes real wiring — calendar, email, context loading — that a webinar template doesn't deliver
• The install tier buys 2.5 hours of calls and 30 days of email support — after that, the 'brain' has no one maintaining it

How to actually do it

• If you're DIY-curious: the stack is real — a Claude subscription plus a structured note vault gets you a genuine starting brain for ~$20/mo, no $99 webinar required
• Before paying anyone for an 'AI brain': ask what BUSINESS CONTEXT gets loaded, what it integrates with, and who maintains it after day 30 — the answers are the product
• Ignore seat counters on evergreen webinars. If it 'sells out' every two weeks forever, it doesn't sell out
• Ask any seller for one named customer running the system 90 days later. No name, no wire

Logged with respect — this is our own category, sold honestly-ish at the concept level. The claims that needed flags got them.

What holds up

• Satire is disclosed in the caption itself — "activate 'ragebait software developers' mode" — no claim is actually being sold
• Creator verified as a real AI builder (~138K followers) whose persona is the product; revenue comes from sponsorships, not viewer purchases
• The underlying premise checks out: documented cases of vibe-coded backends leaking databases and getting hacked within days of launch

What doesn't

• No real "$600k database" story exists — the number is invented for comedy, so don't quote it as a case study
• Ragebait format works precisely because plenty of viewers WILL take it at face value — the disclosure lives in the caption most people never read

How to actually do it

• If you're vibe coding anything that stores customer data: the database config is the part you don't skip — real apps have leaked tens of thousands of IDs exactly this way
• Never ship AI-generated code with API keys in the frontend — it's the first thing attackers check, and it's how the "100% AI-built" SaaS got owned in days
• Enjoy the satire, but check captions before resharing 'claims' — disclosed comedy stripped of its caption becomes someone else's hype

Nothing to build — logged for the kernel, which is a genuinely useful warning wearing a costume.

What holds up

• Repo verified: trycompai/crm on GitHub — MIT license, created July 31, 2026, actively maintained, ~2,535 stars by Aug 3
• README matches the caption nearly word-for-word: 18 tools, 4 markdown skills, evidence-vs-suggestion split, sandboxed egress, durable queue
• Vercel's "eve" durable-agent framework is real (launched June 2026) and is exactly what the architecture leans on
• Comp AI is a real funded company (trycomp.ai, $2.6M pre-seed) that built this for its own sales team first

What doesn't

• "Built on Claude" is the reel's own embellishment — the default agent model in the code isn't Claude (it's swappable)
• The poster built nothing: it's a screen-record of the founder's launch post, riding a 24-hour star spike, with a comment-"CRM" DM gate for engagement
• Days old, ~zero production track record — the README says treat it as promising, not proven. Believe it.

How to actually do it

• Skip the comment-"CRM" DM gate — the repo is public: search GitHub for the Comp AI open-source CRM
• Read the README's design sections first; the no-confidence-scores and sandbox reasoning are worth stealing even if you never deploy it
• Only self-host if you're comfortable with Docker + OAuth setup — this is a weekend project, not a signup button
• Treat it like any 3-day-old dependency: pilot on non-critical data until it has a track record

Logged as a genuine Build-It candidate: the durable-research-agent pattern (agent keeps its own notes, humans confirm weak evidence) rhymes with how we already run our own memory systems.

What holds up

• Real registered agency with a public team page and an active founder LinkedIn
• Three testimonial names verify as real people at real companies
• "Done-for-you funnel + ads in 14 days" is a genuinely deliverable service — the timeline itself isn't the lie

What doesn't

• 3.4x average, 115+ companies, "$95M+ generated": self-reported everywhere, verified nowhere
• Zero third-party review footprint — not even bad reviews; the brand is too new to have a paper trail
• Landing page cloaked from everyone but ad-clickers; guarantee terms published nowhere
• Founder's own posts claim 6–10x on his funnels while the ad averages 3.4x — pick a number

How to actually do it

• Before hiring ANY "ROI guaranteed" agency: get the guarantee's ROI definition, minimum spend, and exit terms in writing before the call
• Ask for 3 client references you pick from their list — not the 3 they volunteer
• Weigh the honest math: a good funnel build is real value; 115-client averages from a team of four with no reviews is ad copy

Nothing to build — this is a vetting-checklist entry. The service category is legitimate; the claims as advertised are unverifiable.

What holds up

• Fanvue's Claude MCP integration is official and documented by Fanvue itself — the reel's core tool claim is true
• The AI video model shown is real and currently available
• AI creators are a real, growing revenue segment on the platform — this niche exists

What doesn't

• Typical earnings: $0–500/month in year one; the headline cases are top-1% agency operations
• Fanvue mandates AI disclosure + ID verification; violations mean removal — the 'passive faceless empire' framing ignores this
• Meta actioned hundreds of thousands of unoriginal-AI-content accounts — your growth channel can vanish
• The poster is unverifiable and the 'free guide' is a comment-to-DM contact harvest

How to actually do it

• If you're genuinely curious, read Fanvue's own MCP and AI-content docs — no guide purchase needed, it's public
• Budget like a business: model the ~20% platform cut, content costs, and months of $0 before judging the hustle
• Disclose AI content everywhere the platform requires it — the account you save is your own
• Skip the DM funnel. Anyone selling a map to a gold rush earns more from maps than gold

Steps given for the honest version; nothing here for us to build.

What holds up

• The download is real — buyers do receive a pack of 70 prompt/skill files
• Claude skills and subagents are a real, useful mechanism — the underlying tech works
• Well-built prompt packs can genuinely speed up marketing tasks

What doesn't

• "Normally $540" in the ad vs. $285 crossed out on his own page — the discount is fiction either way
• A prompt pack styled as a "plugin" with "specialists" — it's text files, and free equivalents (75–100+ subagent repos) sit on GitHub
• Review-site complaints on the seller's other products: overcharging, undelivered access, refund fights
• "Replaces $2–5k/month freelancers" — no prompt file negotiates, strategizes, or owns outcomes

How to actually do it

• Want a team of Claude specialists? Install a free open-source subagent collection from GitHub — several maintained packs ship 75–100+ roles
• Read what's inside any pack before loading it into your agent — free or paid, prompts are dependencies
• If a discount matters to you, screenshot the price from two different devices before buying anything from an info-marketer

Steps given: the free path to everything this ad actually delivers.

What holds up

• Consistent with every real 'AI built this' case we've audited: generation is real, art direction is human
• The poster is a working designer showing work, not selling a course or gating a guide
• No fake numbers, no manufactured urgency, no funnel tricks — a unicorn in this feed

What doesn't

• It's still marketing — the takeaway conveniently equals 'hire a designer'
• No process shown: a before/after of the AI's raw output vs. the directed result would have proven the point instead of asserting it

How to actually do it

• Use AI to build; budget human taste for what customers actually see — that's where trust is won
• When you see a jaw-dropping 'AI made this' demo, ask what the prompt-giver already knew about design
• If you can't afford a designer, steal structure from the best sites in your industry and let AI implement THAT

Logged as honest counter-programming; it's the same lesson our Built-It entries keep teaching.

What holds up

• The prompt/template library exists, is polished, and has a genuinely usable free tier
• The agency behind it has real, verifiable design-industry credentials
• Curated, pre-tested prompts DO cut iteration time versus prompting from scratch — the honest kernel is real

What doesn't

• Undisclosed ownership: the 'helpful tip' is the maker advertising itself in third person
• Their own docs undercut '1:1' — results depend on using the exact model the prompts were tuned for, and exact-match is a paid-tier source-file feature
• 'Cuts 99% of AI garbage' is an invented number with no methodology anywhere
• Zero third-party reviews found; every claim about the product traces back to the product

How to actually do it

• Browse the library directly on its site — no comment-gate needed; the free tier shows you what's real
• If you try a prompt, run it on the exact model their docs specify — that's the honest version of '1:1'
• Never trust a discovery reel without checking who owns the thing being discovered — page transparency takes 30 seconds

Steps given — the free tier is genuinely worth a look for our own site work; the ad's framing is what earned the flags.

What holds up

• Newsletter exists, publishes daily, and is genuinely free
• 125k+ subscribers; documented fast growth in a public beehiiv case study
• Ideas are written editorial content, not scraped filler
• Signup really does take seconds — no credit card, no upsell wall at the door

What doesn't

• 'From a $100M founder' daily is marketing gloss — most ideas historically written by the newsletter team
• The reel is the product's own ad dressed as an observation about 'the new generation of engineers'
• Free tier is the funnel: monetized via sponsor slots and a premium tier

How to actually do it

• Want the ideas? Subscribe at their site — email in, done. It is actually free.
• Want better startup ideas than a newsletter can give you? Mine your own customer complaints and niche communities — ideas you have unfair context to execute beat ideas 125k other people also just read.
• Want what they actually built? Copy the engine, not the ideas: daily short vertical videos with one bold on-screen claim, every one pointing at a single free email capture, sponsors once the list is real. Their beehiiv case study lays out the numbers.

Nothing to build — it's a subscription link, and it's safe. Logged the honest read: subscribe if you want idea flow, but the more valuable move for any small firm is copying their growth engine, which is publicly documented.

What holds up

• greensock/gsap-skills: real, org-owned, "Official AI skills for GSAP," 12.8k+ stars, actively updated
• LottieFiles motion-design-skill: real and genuinely first-party — a taste-and-timing principles library that works with any animation system
• Three.js skill pack: 10 real skills covering scenes, cameras, GLTF, lighting
• Creative-coding plugin: active solo project, shipped a new version within days of the post; slide text matches the README verbatim
• All four use Anthropic’s standard skill format and install cleanly

What doesn't

• "Particle Effects" on the Three.js slide: zero particle content in any of its 10 skills
• The LottieFiles skill won’t touch Lottie/dotLottie files — it teaches motion principles, not export tooling
• The Three.js pack’s own README clone command points at the wrong repository — install by copying its skills folder instead

How to actually do it

• Install the official GSAP skills first — biggest payoff, first-party maintained
• Add the LottieFiles motion-design skill for taste: timing, easing, and choreography guidance on any stack
• Grab the Three.js pack only if you build 3D — and copy the skills folder manually (its own install command is broken)
• Vet third-party skills like any dependency: read what’s in the folder before you load it into your agent

This one goes in the toolbox. Verified worth adopting for real client site work.

What holds up

• All five repos exist under the exact names shown; star counts were accurate within a week’s drift
• The agentic video studio genuinely builds finished faceless shorts via an AI coding agent for $0.15–$3/video in API costs
• The diffusion speedup and talking-avatar tools are legitimate academic-grade projects
• The caption tool does exactly what the slide says

What doesn't

• Zero of the five repos post, schedule, or upload — "posts while you sleep" is covered by nothing in the stack
• Realistic burden: a ~$1,600+ GPU or cloud rental, CUDA dependency builds, 14B-model downloads, and babysitting each render
• The clip-cutter’s entire commit history is one 34-minute day — abandoned, and the wrong tool for this pipeline
• The talking-avatar tool degrades after ~1 minute of video by its own docs
• YouTube demonetizes "inauthentic" mass-produced content; TikTok’s unoriginal-content flag blocks Creator Rewards — the business model dies at the payout step
• The slideshow is top-of-funnel for paid 1:1 calls

How to actually do it

• Treat the stack as a production accelerator: human script and editorial choices, AI does the rendering grunt work
• Start with the caption burner and clip tools on content you actually filmed — lowest cost, immediate payoff
• If you try the generative tools, budget for GPU reality: cloud rental beats buying hardware for a test
• Post manually with AI-content disclosure enabled — monetization policies reward originality, so add your voice, your face, or your take

Used honestly, these tools are a genuinely fast assembly line you operate — not an autopilot.

What holds up

• Content repurposing/clipping is a real, legitimate tactic — one long recording genuinely becomes dozens of clips
• The company exists and some of its view-count screenshots may be real numbers — spam networks do generate views
• Long-form → clips → omnipresence is a sound strategy when run on accounts you own

What doesn't

• Their own public materials describe a mass fan-account network — textbook platform-TOS violation on every network they list
• Anonymous operation: no founder, no address, no legal entity, no terms or privacy policy on the funnel page
• Domain registered Jan 2025 behind a privacy shield; LinkedIn claims "founded 2019" — the timeline doesn’t reconcile
• Zero independent evidence for the Priestley / Brunson / Dawson / MambaFX client claims
• "THEY SHOCKED @MRBEAST" hook: no MrBeast connection exists anywhere in their materials
• Views are not leads — spam-network impressions don’t become customers who trust you

How to actually do it

• Record the long-form content exactly as pitched — that part of the model is right
• Have a small team or an AI-assisted pipeline you own cut it into clips
• Publish on accounts YOU control, where the audience and algorithm equity compound to your brand
• Judge any content vendor by leads and revenue, never raw views — and walk away from anyone whose distribution method can’t survive a platform policy read

The honest version of this offer exists and works — it just doesn’t involve fake accounts.

What holds up

• Jordan Lee is a real person and AI Acquisition is a real company — the persona checks out
• The funnel structure ($97 front end → $5K–$25K coaching tiers) is documented by independent reviewers
• The valuation appears ONLY in his own ads and pay-to-play press releases — no funding database, no priced round, no independent source
• Independent Trustpilot audit: ~92% negative sentiment, incl. "$100K spent, not one cent made"
• FTC has taken repeated federal action against the DFY/automation genre, with bans and eight-figure judgments

What doesn't

• Self-declared valuation as the headline hook — no outside investors means nobody independent ever priced it
• Lifestyle-porn framing: Greece, 4 hours a day, "business runs itself"
• Comments effectively silent on a 23K-view ad — engagement managed
• Classic VSL ladder: free presentation → sales call → five-figure close
• Income claims sourced from a cherry-picked survey of his own customers

How to actually do it

• The honest version: local businesses really do pay for AI implementation — as a service you sell and deliver like any trade. Real work, real revenue, no passivity anywhere
• Before paying anyone to "build you a business": search their name + Trustpilot and read the one-star pages first; then search the genre + FTC
• Valuation test: no outside investors = no independent valuation. Ask who priced it. (There's no answer)
• "Runs itself" test: ask exactly who answers the customers, does the delivery, and handles refunds in month three. If the answer is "AI," ask to watch it do that
• Want the real thing? Pick one deliverable skill — intake bots, memory systems, follow-up automation — deliver it to one local business for a fair fee, and compound. That's the whole secret, and it's free

The kernel is real and worth doing honestly — here's the difference between the fantasy and the actual business, plus the checks to run before anyone pays a guru five figures.

What holds up

• Real device, real funding (seed+angel, ~$M USD), Kickstarter launched April 2026 at ~$300
• Hardware HID + HDMI approach genuinely works with API-less apps
• Physical confirmation key for destructive actions is a legitimately good safety design
• No software install on the host PC; doesn't consume its CPU/GPU
• BYO API key (Claude/OpenAI/Gemini) confirmed on the official site

What doesn't

• '24/7 digital employee' oversells what screen-vision agents reliably do in 2026
• On-device 8B model = routine tasks only; hard tasks hit your cloud API key and bill
• 'Zero detection, zero bans' openly advertises ToS-evasion of platforms that prohibit automation
• Crowdfunded first-gen hardware from a months-old company — delivery and support risk
• No meaningful base of independent long-term user reviews yet

How to actually do it

• Want the outcome now: put Claude Code (or another frontier agent with computer use) on a spare mini PC — same class of automation with a far stronger brain, for the cost of hardware you may already own
• Scope it to tasks with low blast radius: data entry, report pulls, file organization — never irreversible actions unattended
• If Violoop itself appeals, wait for retail availability plus 90 days of independent user reviews before ordering; treat crowdfunded hardware as a preorder gamble, not a purchase
• Whatever runs the automation: check the terms of service of each app you automate — hardware-level 'undetectable' input doesn't change what a platform's rules allow
• Budget the true cost: device price + cloud API usage for any non-trivial task

The honest path to the same outcome exists today without waiting on crowdfunded hardware — laid out below, plus the sane way to evaluate Violoop if the concept appeals.

What holds up

• Retainer-ization of agency services is a proven, legitimate model
• Real company with 700+ reviews; some students report actual retainer wins
• Done-with-you coaching genuinely beats passive course libraries for implementation
• The diagnosis — feast/famine comes from referral dependence — is accurate

What doesn't

• 'AI' appears in the title, not the method — the system described is ads + funnels + scripts
• Price undisclosed anywhere in the funnel; that's a feature of the funnel, not an oversight
• Review wall is heavy on day-one and week-one students, thin on 90-day outcomes
• 'We run this in our own company' is circular proof — their evidence the ads work is that their ad reached you
• Results depend on ad spend the pitch doesn't budget for

How to actually do it

• Repackage your services into 2–3 monthly retainer tiers anchored to a business outcome, not hours or deliverables
• Take it to your existing clients first — the fastest retainer is converting a project client to a monthly engagement at the point of delivery, when trust peaks
• Build one inbound channel before touching paid ads: consistent authority content plus a single clear offer page
• If you evaluate any coaching program: get the total price and refund terms in writing on the first call, and ask for 3 references at YOUR revenue level who are 6+ months out
• Filter their reviews to ones describing results after 90+ days; ignore onboarding-week praise entirely
• Decision rule: pay for coaching only to buy speed and accountability — never information. The information is free.

Extracted the working core — the retainer conversion play — as steps anyone can run without buying anything, plus a vetting protocol if you do want the done-with-you version.

What holds up

• Real company: founded 2023, TinySeed-backed, 500+ agency customers
• Verified users confirm the time savings — 'a day's work in 30 minutes'
• Backlinks come from real journalist/guest-post outreach, placements reviewable
• Users independently confirm DR40+ link placements
• White-label reporting and auto-publish are real shipped features

What doesn't

• '15 minutes' = the setup and plan, not a live campaign — the Demo Cut
• 'Guaranteed' backlinks depend on outreach placement rates that vary by niche
• Users report content accuracy issues and slow external backlink results
• Pricing not public — reviewers peg it around $499–999+/mo behind the 7-day trial
• Scaled guest-post link building lives in Google's gray zone; the ad never mentions it

How to actually do it

• Take the 7-day trial, but connect one real client site — not a test domain
• Generate the campaign plan on day one; that is the genuine 15-minute part
• Publish 2–3 articles to a low-stakes blog and grade them against your current writer's output
• Before trusting 'guaranteed,' ask sales for last month's actual placement rate and 5 sample placements — verify the DR yourself in Ahrefs or a free DR checker
• Get the real monthly price in writing and compare it to your current content + link vendor spend
• Decision rule: if it holds your quality bar and saves 10+ hours/month at that price, keep it; if not, a Claude-based content pipeline plus a vetted link vendor covers the same ground for less

The working core is worth a structured trial for any agency doing SEO delivery. Wrote the exact evaluation path below — decision rule included, so the trial ends in a keep-or-kill call, not a vibe.

What holds up

• Obsidian vault = plain local markdown; Claude reads and writes it directly
• Claude Code needs zero plugins — point it at the vault folder
• Independently reproduced by multiple writers, Feb–Jun 2026
• Notes stay local; nothing uploaded beyond normal Claude usage
• Setup consistently reported at 20–30 minutes

What doesn't

• Silent video — the exact on-screen method can't be verified from outside the platform
• A second brain is only as smart as the capture habit feeding it — the demo skips that part

What we did

• Install Obsidian (free, obsidian.md) and create a new vault — one folder of markdown files
• Inside the vault create folders: Inbox, Daily, Projects, People, Resources, Maps, Templates
• Create CLAUDE.md at the vault root describing the folder rules and how you want notes filed, linked, and summarized — this becomes Claude's operating manual
• Install Claude Code (docs.claude.com), open a terminal, cd into the vault folder, run: claude
• First prompt: 'Read CLAUDE.md, index every note, and write Maps/overview.md linking the major themes'
• Prefer an app over a terminal? Claude Desktop → Settings → Extensions → Filesystem → add the vault folder — same result
• Daily use: dump raw thoughts into Inbox, then ask Claude to file, link, and summarize them per CLAUDE.md

Verified the full stack against current docs and independent reproductions, then built a ready-to-drop-in vault starter kit (folder structure, CLAUDE.md operating manual, capture templates). The steps below work standalone with no kit and no course.

What holds up

• Interview-first drafting works in any capable chat AI today, free tier included
• Google's own docs reward original info and firsthand experience
• Voice-mode answers are a real, low-friction way to capture expertise
• One-question-at-a-time pacing produces far better raw material than a single dump

What doesn't

• "99% more likely to rank" — fabricated statistic, sourced from nowhere
• "100x easier" — unfalsifiable hype math
• E-E-A-T guidelines inform quality raters; they are not a direct ranking dial
• Prompt skips keyword research, search intent, internal links, technical SEO entirely
• Reframed basic: this is standard journalism interviewing, rebranded as an AI hack

What we did

• Pick one real question a customer has actually asked you — that's the post topic
• Paste the upgraded prompt: "I'm writing a blog post answering: [question]. First ask me the target search phrase and who's searching it. Then interview me — max 10 questions, one at a time — digging for firsthand stories, numbers, mistakes, and opinions a competitor couldn't fake. I'll answer by voice; wait for each answer. Then draft the post: answer the question in the first 100 words, use my exact stories, add H2s phrased as related questions people search, and end with 3 internal-link suggestions and a meta description under 155 characters."
• Answer out loud with specifics — client names redacted, numbers kept; specifics are the whole value
• Edit the draft for voice, add one original image or screenshot (experience signal Google explicitly looks for)
• Before publishing, verify the target phrase actually gets searched — free check in Google Keyword Planner or by typing it into Google and reading the autocomplete

Extracted the working core and upgraded it: an interview-to-post workflow that also bakes in the SEO layer the original skips (target query, search intent, headers that match the question, internal-link prompts). Full upgraded prompt in the steps — works in Claude or ChatGPT.

What holds up

• Claude ↔ Obsidian works today — Desktop filesystem connector or Claude Code in the vault folder
• Vault is plain markdown you own; no lock-in, works offline
• One-time setup is honest — roughly 15 minutes
• The pattern is replicated by many independent creators, several with truly free templates

What doesn't

• “Run your life” = reads and writes notes; nothing runs unattended out of the box
• It's a paid ad (‘Learn more’ funnel) — the template exists to capture you, not to help you
• Claude–Obsidian integration is under two years old; a ‘3-year’ Claude-connected template can't exist as framed
• Documented buyer complaints in this genre: ‘free template’ gates that never deliver the files

What we did

• Install Obsidian free from obsidian.md — ‘Open folder as vault’ on any folder
• Route A: Claude Desktop → Settings → Connectors → Filesystem → add the vault folder
• Route B: terminal in the vault folder → type ‘claude’ (Claude Code reads CLAUDE.md automatically)
• Add a CLAUDE.md telling it your folder map and standing behaviors — this is the whole ‘template’
• First prompt: ‘Read CLAUDE.md, interview me, and build out my projects, clients, and people notes’

Built the template they're selling and gave it away: a Claude-ready Obsidian starter vault (CLAUDE.md operating instructions, daily/client/project/person templates, inbox workflow) plus the exact connection steps.

What holds up

• AI-drafted ads, nurture sequences, and auto-booking are standard, working tech
• Pre-qualifying before a human call genuinely raises close rates
• Keeping the human on the sales call is honest — full-AI closing is worse
• 2-3 hrs/week of oversight vs 20 hrs of manual prospecting is a fair before/after

What doesn't

• "Genuine ads" = paid ads; the required monthly ad spend is never stated
• "Stages 1-4 fully automatic, no involvement" ignores creative testing and offer work
• Automated SMS at scale requires A2P 10DLC registration and TCPA consent — never mentioned
• "Booked while you sleep" depends on the offer and ad budget, not the plumbing

How to actually do it

• CRM + automation platform with pipelines, email/SMS, and calendar booking built in (GoHighLevel-class tool, roughly $100-300/mo)
• Register your business for A2P 10DLC before any automated SMS — carrier registration plus documented opt-in consent; skipping this risks TCPA liability
• Draft the ad angles and 14-touch nurture sequence with Claude, in your voice, against your real offer
• Meta ads with a realistic test budget — expect four figures monthly before the calendar looks like the demo
• Human takes every sales call; review the sequence's reply rates weekly and rewrite the weakest touch

Broke down the honest stack, real running costs, and the compliance step the reel skips.

What holds up

• Anthropic's agent guidance is real, public, and genuinely excellent
• The principles named (simplicity, tool design, evaluation) do appear in the essays
• The agent loop — gather, act, verify, repeat — is the actual documented model

What doesn't

• "$2.3M" appears nowhere; the documents are free blog posts
• "4 years of development, 1000+ hours of lectures" — fabricated dressing
• No single "loop manual" exists; it's separate essays, synthesized by the community
• "An agent that earns" — the docs teach reliability, not revenue

What holds up

• Dedicated single-tenant VM is real — persistent disk, daily backups, 24/7 scheduled jobs
• Established vendor: 2,100+ teams, $14B/yr in ad spend analyzed
• Named customers with specific use cases, not vague testimonials
• Real pricing page, unlimited seats, 30-day money-back guarantee
• Their own FAQ admits you "can get partway" with general AI tools

What doesn't

• "Every tool in your stack" = a fixed list of ~40 integrations
• The ad contains zero pricing — real entry is $750/mo plus $500/mo usage packs
• "100X marketer" is pure puffery, no measurable outcome claimed
• No free trial — money-back guarantee stands in for one
• Their comparison table calls Claude a "single-player chat that can't touch your live data" — stale; MCP connectors do exactly that today
• First-party ad: the poster is the seller

How to actually do it

• Spending $10k+/mo on paid social and want their creative dataset? Activate the Starter tier ($750/mo, unlimited seats) and treat the 30-day money-back guarantee as your trial — write a list of deliverables it must ship in month one, and refund if it misses.
• Otherwise, build the same shape on Claude: a Claude Pro/Max plan ($20–200/mo) with Cowork or Claude Code on a machine that stays on.
• Connect the same stack via first-party MCP connectors in Claude's settings: Slack, Google Drive, Notion, Canva, and your PM tool. Point-and-click, no code.
• Recreate the "brain": a folder of markdown memory files (brand voice, naming conventions, offers, audience notes) plus reusable skills for repeat workflows — briefs, reports, ad QA. The demo calls this "training"; it is writing things down once.
• The one piece you cannot DIY is their dataset of millions of tagged ads and benchmarks. If that is the draw, that is what the $750/mo is for — price it against one month of a junior analyst, not against the AI.

Split the path: if the tagged-ads dataset matters and your ad spend justifies it, the paid product with its 30-day guarantee is the honest route. Otherwise the same architecture self-assembles on Claude for roughly a tenth of the cost — exact steps below.

What holds up

• AI genuinely drafts a book-length guide in a day
• The prompt "works" — because any outline prompt works
• Reproduced it live: finished 10-page branded guide in one session

What doesn't

• The "secret prompt" is a generic one-liner, free anywhere
• Outline ≠ product: audience, offer, trust, distribution all skipped
• 2026 data: AI titles = big share of listings, small share of sales
• Amazon KDP requires AI disclosure; bans mass-produced slop
• Prompt-pack funnel at the link in bio

What holds up

• The workflow is real — it's the industry-standard 4-phase process, fully disclosed in the caption itself
• Each phase description is accurate, ordinary professional practice
• Not strictly an AI claim — logged because the gate mechanics are the same machinery as the AI funnels in this audit

What doesn't

• The gated "secret" is 100% visible above the gate — the comment buys you nothing you haven't already read
• "Websites of this quality" is unverified — this reel genre is notorious for showcasing other people's award sites
• "Exact workflow" framing manufactures scarcity around the most generic process in the industry
• The gate's real function: comment volume feeds the algorithm and fills a DM pipeline

How to actually do it

• Don't comment. You already have the workflow — it's the caption. Here's the operator version of each phase:
• Direction: one page answering who it's for, the one action a visitor takes, and 3 reference sites — this page prevents every expensive revision later
• Design: wireframe before pixels; lock typography and spacing as reusable styles so the build doesn't drift
• Build: develop mobile-first, add motion last — motion is seasoning, not structure; AI tools (Claude, v0) now compress this phase from weeks to evenings
• Delivery: test on a real phone, run PageSpeed, hand off hosting access and an edit guide — handoff quality is what gets referrals
• If you run an agency: notice the tactic itself — gating works even with nothing behind it; the ethical version is gating a genuinely extra asset and saying plainly what it is

Confirmed the gated workflow against the caption it's printed in, and wrote out the operator version of the four phases — the part a DM PDF would actually skip.

What holds up

• AI-generated 3D web scenes are real — Claude and peers write working Three.js from a plain description
• "An evening instead of weeks" holds for a working draft scene
• The essay's core (AI collapses the cost of personal projects) is genuinely true — same as this page's entry #004

What doesn't

• "1:1" from a prompt is still mechanically false on the prompt tier — LLM output varies per run
• The 1:1 experience requires the top Full Stack tier, where you're buying source code, not prompts
• "Cuts out 99% of AI garbage" is an unverifiable marketing number
• The reels run faceless with double gates while the company site is public — the ads hide what the product discloses
• Interchangeable philosophy essays around an identical sales block across multiple reels — a template farm

How to actually do it

• For guaranteed-identical results, skip prompts entirely: threejs.org/examples has hundreds of official working scenes with source — copy, run, restyle
• For production-grade showpieces, Codrops (tympanus.net/codrops) publishes free 3D and interaction demos with complete downloadable code
• For custom scenes, describe what you want to Claude like a brief: "a slowly rotating wireframe globe with glowing city markers, dark background, mouse parallax" — expect a working draft in one pass and 3–5 refinement prompts to taste
• In Claude artifacts specifically, ask for Three.js r128 — it's the supported version and renders live in the chat
• If you do want the curated route: GetLayers has free sample prompts and a real Claude Code plugin (Full Stack tier) — evaluate it at getlayers.ai directly instead of DM-ing a gate word
• Budget honestly: one evening to a working scene, a second session to make it client-polished — same arithmetic as this page's own essay
• Never DM a gate word to an anonymous page for access to a platform it won't name in public

Originally graded on mechanism alone; a follow-up reel identified the platform, and inspection of its site confirmed the two-tier resolution predicted here — verdict upgraded to High confidence.

What holds up

• $47B run-rate revenue: really disclosed by Anthropic at its May 2026 Series H
• The publication exists and is free: Anthropic's multi-agent research system engineering post
• The architecture is described accurately: lead agent plans, parallel subagents execute
• The published numbers are real: 90.2% over single-agent baseline, ~15x token cost

What doesn't

• "Graph Engineering" is not Anthropic's term — it's a paid community's rebrand of a free blog post
• Coordinated wave: multiple accounts, same week, same "Anthropic just dropped" template, same destination
• "Shouldn't have made this free" manufactures leak-energy around a post that's been public since 2025
• Run rate ≠ annual revenue — the real number, stretched one notch for the hook
• The omitted cost: the multi-agent pattern burns ~15x tokens, which the thread never mentions

How to actually do it

• Read the actual source, free: Anthropic's engineering post "How we built our multi-agent research system" — the lead-agent-plus-subagents pattern, with the real numbers and the real tradeoffs
• Then the docs the threads paraphrase: Claude Code's subagents — markdown files in .claude/agents/, each with its own context window, system prompt, and scoped tools
• Try it in any Claude Code project: /agents creates one interactively; a researcher + a writer + a reviewer is the classic starter trio
• Respect the cost line Anthropic itself publishes: multi-agent runs ~15x the tokens of a chat turn — use it where parallel work pays for itself, not for tasks one agent handles fine
• Skip anything charging to teach "Graph Engineering" — the sources above are the curriculum, verbatim and free

Traced the term to its actual coiners, verified the revenue figure and the architecture against Anthropic's own publications, and pointed at the free originals the threads are reselling.

What holds up

• Blotato MCP is real: one endpoint, 9 platforms, publish + schedule from a Claude prompt
• Works in Claude Code, Claude Desktop, and claude.ai as a connector — no Meta API setup
• The caption honestly includes the confirmation-before-posting step
• "This is what MCP does" — the one-line MCP explanation is actually correct

What doesn't

• "Free access" — Blotato has no free plan; $29/mo minimum, so the DM delivers a trial
• Undisclosed in the caption: the poster founded the company the reel recommends
• Adds a paid Claude plan on top — two subscriptions behind a "free" hook
• Instagram publishing requires a Business or Creator account, not a personal profile
• The DM-"POTATO" gate is the same comment-to-funnel machinery as her other reels

How to actually do it

• Decide the loop you want: drafts-for-approval only (free) or approved-drafts-that-publish-themselves (paid layer)
• Free route: Buffer's hosted MCP (mcp.buffer.com/mcp) on Buffer's free plan — solid for 3 channels and basic queue scheduling
• Paid route: Blotato Starter at $29/mo is legitimately the strongest publish layer if you post video across TikTok, YouTube, and Instagram — one MCP, 9 platforms, no Meta developer-account pain. Sign up at blotato.com directly; the POTATO DM adds nothing but a funnel entry
• Connect in Claude Code with the /mcp command per the setup docs at help.blotato.com/api/mcp, then authenticate in the browser
• Keep the approval gate: prompt Claude to always show drafts and wait for an explicit yes before any publish call — never run it fire-and-forget
• Convert Instagram to a Business or Creator account first if it's still personal — Meta blocks API publishing to personal profiles

Verified the MCP against its live docs and priced the honest version — including when the paid tool is genuinely the right call versus the free route.

What holds up

• Prompt-to-website is a real, mature capability: Claude, v0, Lovable, Bolt all deliver
• "The price of an attempt is one evening" is genuinely true — and worth internalizing
• The caption is honest that this promotes a project, even if it never names it

What doesn't

• No evidence the gated prompts produced the sites shown — this format is cherry-pick central
• Anonymous page, unnamed "project" — you're DM-ing your engagement to a funnel with no face
• Comment gates exist to train the algorithm and harvest DMs, not to share knowledge
• One "master prompt" doesn't replicate a polished site — real results take iteration passes
• Whatever the link sells arrives only after you're inside the DM automation

How to actually do it

• Skip the gate. Open Claude (claude.ai, Pro $20/mo) and describe the site like a creative brief, not a wish: business name, who it's for, the one action a visitor should take, 3 sections you want, and 2 example sites whose style you like
• Ask for a single-file HTML landing page first — one file is easy to preview, host, and iterate on
• Iterate in plain English 3–5 times: "darker background", "make the headline bigger", "add a testimonial section" — iteration, not the first prompt, is where the polish comes from
• Free hosting: drag the HTML file into Netlify Drop (app.netlify.com/drop) — live URL in under a minute, $0
• If you want the heavier route: v0.dev or lovable.dev generate full React sites from the same kind of brief — free tiers exist, paid from ~$20/mo
• Total honest budget: one evening for a real draft, a second evening to make it good

Graded the checkable claims, flagged the unverifiable one, and handed over the ungated version: how to get a real website out of one evening with AI, no comment required.

What holds up

• The build is real: Claude Code skills + brand voice + quality-gate hooks, fully documented
• One-sentence-in, week-of-drafts-out genuinely works once the system exists
• The complete tutorial with exact prompts is free on the poster's blog — no email gate
• Poster is honest elsewhere about reviewing every piece before it goes live

What doesn't

• "1 minute" excludes the ~90-minute build the poster's own tutorial prescribes
• "Finished posts" skips the human review step baked into the poster's own workflow
• The publish pipeline runs through the poster's own SaaS — $29–499/mo, no free plan
• Needs a paid Claude plan on top — the post prices the whole thing at zero
• Brand voice quality depends on feeding it your real writing samples — garbage in, generic out

How to actually do it

• Install Claude Code (claude.com/claude-code) — a Pro plan ($20/mo) covers a solo content workload
• Follow the free tutorial at sabrina.dev/p/claude-code-full-course-for-beginners — budget 90 minutes, not 1
• When it asks for brand voice samples, paste 5–10 of your real posts — this single step decides whether the output sounds like you or like AI slop
• Keep the quality-gate hook step — it's the part that catches em-dash-riddled generic output before you ever see it
• For scheduling, skip the $29/mo SaaS: Buffer runs a free hosted MCP server at mcp.buffer.com/mcp with published Claude Code setup guides, and its free plan handles basic queue scheduling
• Run it as drafts-for-approval, not fire-and-forget — the creator herself reviews every post, and that's the honest version of this system

Confirmed the build against the poster's own free 90-minute tutorial, priced the hidden subscriptions, and handed over the setup — including a free scheduling route that skips the paid SaaS entirely.

What holds up

• Remotion ships official Claude Code skills: npx skills add remotion-dev/skills
• Renders locally — no credits, no per-video fees; video 1 costs the same as video 100
• Independent reproduction: 2 prompts, ~15 minutes, usable branded 30-second video
• Remotion's free license allows commercial use for individuals and companies up to 3 people

What doesn't

• "Near $0" hides the engine: a paid Claude plan, with heavy token burn per video build
• "1,000 videos" shown nowhere — every usable clip still needs iteration prompts
• First drafts are rough; reproducer's own words: "not beautiful, but pretty damn good"
• For-profit companies with 4+ people need a Company License — $25/seat, $100/mo minimum
• The comment-gated "5 master prompts" are already published free on the poster's blog

How to actually do it

• Install Claude Code if you don't have it (claude.com/claude-code) — any paid Claude plan works; Pro ($20/mo) is enough to start
• In your terminal run: npx skills add remotion-dev/skills — this installs Remotion's official agent skills
• Run: claude — then prompt: "Use the Remotion best practices skill. Create an educational explainer video (1080x1920, 30fps, 30 seconds) that teaches '<your topic>'. Keep all text 150px from the top and 170px from the bottom."
• Skip the comment gate — the 5 ready-to-paste prompts are free at sabrina.dev/p/5-insane-claude-code-video-prompts (explainer, product demo, testimonial showcase, avatar edit, data viz)
• Preview in Remotion Studio (Claude opens it at a localhost URL) and iterate in plain English: "slow the typing", "change accent color", "make the bars wider"
• When it looks right, tell Claude to render the MP4 — output lands in the project folder, ready to post

Verified the workflow against Remotion's official skills, licensing, and independent reproductions, then handed over the exact no-funnel setup — including where the "gated" prompts live for free.

What holds up

• Repo is real, free, no email gate — 295 stars, v3.2 released July 2026
• Markdown-as-memory is a proven pattern used far beyond this creator
• Costs disclosed up front: paid Claude plan required, everything else free
• "No coding" holds — Claude runs the build script and interviews you

What doesn't

• "Never forgets" oversells file storage — recall depends on priming and vault upkeep
• "Ten minutes" skips Windows PATH friction, login, and Obsidian setup — plan an hour
• CC BY-NC-SA license: noncommercial — repackaging it for paying clients needs permission

How to actually do it

• Install Claude Code: in PowerShell run irm https://claude.ai/install.ps1 | iex — then add %USERPROFILE%\.local\bin to your user PATH and reopen the terminal
• Install Obsidian (free, obsidian.md) and let it create a vault folder
• Download ai-memory-vault.md from the ai-memory-vault repo on GitHub into your project folder
• In that folder run claude and say: "Read ai-memory-vault.md and build my memory vault." It interviews you and generates CLAUDE.md, VAULT-INDEX.md, and the daily-note template
• Keep CLAUDE.md in the project folder, the vault in Obsidian, and start every session from that folder so the agent boots with its memory

Verified the method against the repo and independent implementations of the same pattern, flagged the noncommercial license, and handed over exact finish-line install steps.