A 60-line rules file, built from Andrej Karpathy's complaints about AI coding agents. Add it once and Claude stops making silent assumptions, overcomplicating everything, and editing code you never asked it to touch.
What this actually is. Andrej Karpathy, a founding member of OpenAI, pointed out the things AI coding agents do that drive him insane: they assume instead of asking, they overcomplicate everything, they change code you never touched, and they call a job done without checking it. Someone turned those complaints into a short set of rules. You add it to Claude Code, and it reads them at the start of every session. That's the whole thing.
Open Claude Code, then type the command below and hit enter. This tells Claude Code where these rules live so it can install them.
/plugin marketplace add multica-ai/andrej-karpathy-skills
One more command. This loads the Karpathy rules into Claude Code.
/plugin install andrej-karpathy-skills@karpathy-skills
Type /exit to close Claude Code, then start it again with claude. From now on it reads these rules at the start of every session. You don't have to do anything else.
claude
The change is in how it behaves, not what it says. Give it one of these and watch what it does differently.
Add user login to my app.
Deliberately vague. Before, it would invent a whole auth system. Now it should ask what you actually want first.
Fix this bug. Change as little as possible and explain what you touched.
Before you build this, tell me what done looks like and how you will check it.