What Andrej Karpathy Skills is
Andrej Karpathy Skills is a compact repository whose value is a single `CLAUDE.md` file. The file gives Claude Code stricter working rules around common LLM coding failures: hidden assumptions, overengineering, side-effect edits, and weak verification.
It is not an official Karpathy project and not an agent framework. It is a public prompt-engineering artifact that can be placed in a project so the agent names uncertainty, keeps changes narrow, and finishes with a verifiable result.
What is inside
The file is organized around four principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. Each principle maps to a real failure mode in coding agents.
Coding-agent rule shape
This mirrors the repository: not a library API, but a project instruction file that changes agent behavior.
# Claude Code project rules
- State assumptions before editing
- Prefer the smallest working change
- Touch only files needed for the task
- Run or explain verification before finishingWhy people use it
Repositories like this became useful as Claude Code, Codex, Cursor, and similar tools moved into real codebases. The model may solve a local task yet still damage a worktree through confidence or unnecessary breadth. A small rules file puts engineering discipline close to the agent.
Strengths and limits
Its strength is minimalism: one file is easy to read, edit, and test in a workflow. Its limit is also clear: it does not replace tests, review, CI, or permissions. It shapes behavior; it does not make an agent automatically reliable.