multica-ai/andrej-karpathy-skills
multica-ai/andrej-karpathy-skills
What it is
multica-ai/andrej-karpathy-skills is a GitHub repository containing a compact CLAUDE.md guideline file for Claude Code / coding agents. The repository description says it is derived from Andrej Karpathy's observations on common LLM coding pitfalls, but it is maintained by multica-ai, not by Karpathy directly.
For Karpathy-X-Radar, the durable value is not the repository's popularity by itself; it is the way it turns social observations about agent failure modes into an executable project instruction file. It is therefore best read as a near-primary synthesis source for External-Agent-Skills-Design-Patterns and Agent-Skill-Discovery.
Four principles
The raw CLAUDE.md organizes the guidance into four operational principles:
- Think Before Coding — do not silently assume; surface uncertainty, tradeoffs, and simpler alternatives.
- Simplicity First — write the minimum code that solves the task; avoid speculative abstractions and unnecessary configurability.
- Surgical Changes — touch only what the request requires; do not refactor adjacent code or clean unrelated dead code.
- Goal-Driven Execution — turn tasks into verifiable goals, ideally with tests or checks before claiming completion.
Why it matters
The page is a useful bridge between AI-Self-Improvement-Lab and day-to-day coding-agent operation. It suggests a pattern: instead of asking an agent to “be better,” encode recurring failure modes as explicit, testable rules in a local instruction file or skill.
For Hermes skill design, the strongest reusable pieces are:
- Assumption management: name ambiguity before making irreversible changes.
- Diff discipline: every changed line should trace to the user's request.
- Anti-bloat bias: prefer smaller implementations unless requirements justify abstraction.
- Verification loop: define success criteria, run checks, and report real tool output.
Caveats
This is not a primary Karpathy repository. It should not be cited as “Karpathy's official CLAUDE.md.” Treat it as a community implementation inspired by Karpathy's public comments. Claims about measured error reductions from secondary media should be verified separately before being promoted to high-confidence wiki knowledge.