Andrej Karpathy Skills
I drop andrej-karpathy-skills into Claude Code because it forces the model to honor repo style, flag dead code without deleting, and push back when a simpler approach exists, Karpathy's 4 LLM-pitfall rules in one CLAUDE.md. Andrej Karpathy (OpenAI co-founder) noticed every LLM makes the same mistakes: over-engineering, unrequested refactors, useless dependencies. A developer turned his observations into 4 rules in a single file.
Strengths
- 4 simple rules covering the 3 recurring LLM agent-mode failure modes
- Born from Karpathy observations (OpenAI co-founder), strong technical authority
- Distributed as a single CLAUDE.md file, trivial install
- 60K stars in a short time, heavy community adoption
Limitations
- Biases toward caution, unnecessary overhead on trivial tasks
- One rule ("contradict the user") can friction with fast tactical workflows
Best for
- Serious refactors where LLM over-engineering costs more than the feature to ship
- Codebases with strong style that the model must respect against its own instincts
- Teams wanting a shared LLM frame to drop into every repo