Four rules from Karpathy that fix most agent coding mistakes

andrej-karpathy-skills is a single CLAUDE.md built from Andrej Karpathy's posts about where coding models go wrong: unchecked assumptions, overcomplicated code, and edits to things nobody asked to change. It boils that down to four rules.

The rules

• Think before coding. State your assumptions, and when a request can mean two things, say so instead of silently picking one.

• Simplicity first. The minimum code that solves the problem. No speculative features or one-use abstractions.

• Surgical changes. Touch only what the task needs, match the existing style, and mention dead code rather than deleting it.

• Goal-driven execution. Turn the task into something checkable, such as a failing test, then make it pass.

Install

As a Claude Code plugin:

/plugin marketplace add forrestchang/andrej-karpathy-skills
/plugin install andrej-karpathy-skills@karpathy-skills

Or drop the file into one project:

curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md

There's a Cursor rule in the repo too.

Why it works

The last rule does most of the work. An agent with a clear finish line, like a test that has to pass, stops wandering. If you only adopt one habit, give every task a way to prove it's done.