Karpathy named the failure modes in late 2025. Chang turned them into a 70-line file (Jan 2026) that crossed 110k stars — one of the fastest-growing GitHub repos of the year. The artifact matters more than the prose: a copy-pasteable boilerplate clients will already have heard of.
What's novel beyond the four rules: a tech-stack lock, mandatory change-summaries, hard stops for destructive/production actions, and MEMORY.md / ERRORS.md — persistent decision and failure logs. That last pair is the genuine contribution — a concrete instantiation of the Write operation in context engineering.
Don't repeat the "65% → 94% accuracy" number to clients. No methodology was published. The accompanying "$975/dev/week" math is napkin-quality. Citing either would undermine our credibility. See Synthesis · honest outcomes.
Maps onto patterns we already have:
- Think-before-code · simplicity-first → Plan mode + workflow ladder.
- Surgical changes · destructive-action confirmation → tool tiering.
- Flag uncertainty → Jaggedness + verifier patterns.
- MEMORY.md / ERRORS.md → the Write operation in context engineering.
Karpathy (observation) · Forrest Chang (distillation) · multica-ai/andrej-karpathy-skills
Recommend the file as a starting boilerplate — tailored to client stack/voice, percentage stripped. Cautionary tale for the prompt-file linter: don't cargo-cult all 70 lines.