← The AI Hype Audit — all 244 verdicts
PARTLY "Claude can build this, but it's never gonna design this" — a fair point wearing an unfair word: 'never'
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.
What holds up
- All named repos are real: Superpowers (obra), RAGFlow (infiniflow), Karpathy’s nn-zero-to-hero/nanochat, and the awesome-claude collections
- Karpathy’s repos are legitimately superb free education — 120K+ stars across his nano projects
- Superpowers and RAGFlow are serious, maintained tools
What doesn't
- 'Never' is unsupportable — current Claude produces coherent, sometimes distinctive visual design today; the gap is context, not capability
- The slick site used to 'prove' human superiority is exactly what AI produces unassisted — and viewers can't verify it wasn't AI-assisted
- Soft self-interest: it's authority content from a designer whose livelihood the claim defends
- The build-vs-design binary is already stale — the real 2026 split is AI defaults vs human-directed AI output, and the winners use both
The catch
Real repos, borrowed authority: the list works as a bookmark folder and fails as a curriculum — and its most impressive number is the one already debunked.
How to actually do it
- Pick ONE — Karpathy’s nn-zero-to-hero — and actually build through it end to end
- Install Superpowers only when you’re already using Claude Code daily; it’s a workflow amplifier, not a course
- Before trusting any repo’s benchmark, search for independent replications — Caveman’s gap (65% claimed, ~9–14% measured) is the cautionary tale
Six real repos, two of them just link lists, one with a debunked stat — one repo built-through beats six starred.
- 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
We test hype for free. We build the real thing for a living.
Thirty minutes, no pitch — and you'll leave with something useful either way.
Book a call with Todd or start with the free Business Checkup →The Verdict Weekly
Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.
© Schreier Group · schreiergroup.com · See a wild AI claim? Drop it here and we'll test it.