Andrej Karpathy's rules for Claude Code, in one file
Four principles in one file. Drop it into a repository and the agent stops guessing and stops overbuilding.
This is a single CLAUDE.md guidelines file, derived from Andrej Karpathy's publicly stated advice on working with AI coding agents, that you copy into a project to change how Claude Code behaves. It encodes four principles: think before acting, keep the solution simple, make surgical rather than sweeping edits, and work goal-first with tests before code. Its main effect is to stop the agent making silent assumptions and running with them, and to stop it rewriting more than the task needs. There is nothing to install and no dependency: you copy one Markdown file into the repository root.
What you get
- One file that changes how a coding agent behaves in your repository
- Four principles: think first, keep it simple, surgical edits, goal-driven with tests first
- A direct fix for agents that assume instead of asking
- A curb on agents that rewrite half a file for a one-line change
- No install, no dependency, no runtime
Who this is not for
It is a guidelines file, not a tool. It cannot enforce anything: it steers a model that may still ignore it. If you want enforcement, hooks and verification, a full harness such as ECC is the different thing you actually want.
Credit
Created by multica-ai· Derived from Andrej Karpathy's published guidance
ProfitizeOS did not build this and does not host it. You get a link straight to the original.
Free. Two fields, then straight to the resource and setup guide.
Common questions
Did Andrej Karpathy write this repository?
No. It is a community-assembled CLAUDE.md derived from his publicly stated guidance on working with coding agents. Credit the repository author for the file and Karpathy for the underlying advice.
Is there anything to install?
No. It is one Markdown file. You copy it into a project and the agent reads it.
Does it work with agents other than Claude Code?
The file is named and formatted for Claude Code. The principles transfer to other agents, but you would need to place the content wherever that agent reads its project instructions.
Is it licensed for commercial use?
The repository does not ship a standard licence file, so there is no clear grant. For commercial work, the safest path is to read the principles and write your own equivalent rules.
Read first
More on this topic: AI Automation