forrestchang/andrej-karpathy-skills

This project is a single configuration file called CLAUDE.md, designed to make Claude Code (an AI coding assistant) behave better when helping developers write software. It is based on observations from Andrej Karpathy about common pitfalls that large language models fall into when generating code. The problems it tries to fix are familiar to anyone who has worked with AI coding tools: the model makes silent assumptions, writes more code than necessary, edits files it was not asked to touch, and removes or changes things it does not fully understand. The solution is a short rulebook organized around four principles. Think Before Coding asks the model to surface its assumptions and ask questions when ambiguous. Simplicity First tells it to write the minimum code that solves the problem and avoid speculative abstractions. Surgical Changes restricts edits to only what was requested, even if nearby code looks improvable. Goal-Driven Execution reframes tasks as verifiable success criteria, often by writing tests first, so the assistant can loop until it gets things right. You would use it if you regularly work with Claude Code (or, with the included rule file, Cursor) and find the assistant overcomplicates things or makes unwanted changes. The repository is distributed either as a Claude Code plugin or as a CLAUDE.md file you copy into a project. There is also a Chinese-language version. The project is MIT-licensed.