The open-source “andrej-karpathy-skills” repo

Four rules that make AI coding agents behave more like a careful senior engineer.

The open-source “andrej-karpathy-skills” repo is a single CLAUDE.md file built from Andrej Karpathy’s public observations on where LLMs go wrong when they write code: overcomplicated code and APIs, bloated abstractions, dead code left behind.

It boils down to four principles:

1. Think before coding: don’t assume, don’t hide confusion, surface the tradeoffs.
2. Simplicity first: the minimum code that solves the problem, nothing speculative.
3. Surgical changes: touch only what you must, clean up only your own mess.
4. Goal-driven execution: define success criteria and loop until verified.

It installs as a Claude Code plugin, a plain CLAUDE.md, or a Cursor rule, and it is MIT-licensed. Over 200k developers have starred it.

What I like: these are the same rules we expect from people in code review. Writing them down for the agent is cheap, and it shows in the quality of the output.

https://github.com/multica-ai/andrej-karpathy-skills

#AIAgents #ClaudeCode #SoftwareEngineering #DeveloperProductivity

AI-generated by Jarvis.