Karpathy's programming experience, Skills, has been open-sourced, and its stars are skyrocketing. It's highly recommended to feed your AI with this andrej-karpathy-skills file. This project is simple: it compiles Karpathy's criticisms of poor coding practices for large models into constraint instructions that large models can understand. This file, less than 70 lines long, has garnered nearly 60,000 stars. It originated from Karpathy's previous observations of common AI programming problems: guesswork, over-engineering, and haphazardly modifying unrelated code. Developer Forrest Chang condensed these experiences into four core rules: 1. Think before you write: Ask questions before making assumptions when encountering ambiguity. 2. Simplicity first: Don't add unnecessary features; avoid over-design. 3. Precise modification: Only modify what needs changing; don't touch surrounding code, no matter how messy it may be. 4. Goal-driven: Give the AI clear success criteria (e.g., passing tests), rather than vague instructions. Download this file to your project's root directory as... Simply use the command `CLAUDE.md` or `AGENTS.md` to let the AI read the code; after that, its work will become much more efficient. Claude Code users can also install the plugin globally with a single command. AI code writing speed is indeed fast, but these four principles are like holding its reins. Knowing when to let the AI run and when to give it a helping hand can help you avoid many hidden pitfalls.