The Karpathy CLAUDE.md: One File, Four Rules, 100k+ Stars

Fifth in the series is the opposite of everything else covered so far: no scripts, no plugin architecture, just one file. andrej-karpathy-skills, built by developer Forrest Chang, is a single CLAUDE.md that has picked up well over 100,000 stars.

Where it came from

In late January 2026, Andrej Karpathy posted about the recurring failure modes he was hitting after shifting most of his coding work over to agents: models making silent wrong assumptions and running with them, bloating simple fixes into sprawling abstractions, and quietly touching code that was never part of the request. Karpathy didn’t write any tooling around this himself – Chang built the file that addresses the three problems directly, and Karpathy hasn’t publicly endorsed the result.

Four rules, not a framework

The file encodes four behavioural principles: don’t make silent assumptions, don’t over-engineer past what the task needs, don’t make orthogonal changes to code outside the request, and define explicit, verifiable success criteria up front rather than declaring victory on vibes. It’s a behavioural file that’s always active for a project, not a task-triggered SKILL.md.

It’s worth being clear-eyed about the trade-off: because it pushes Claude to surface assumptions and ask before proceeding, it adds friction on genuinely trivial one-line fixes where that clarification isn’t needed. It’s aimed at non-trivial work where a wrong silent assumption is expensive to unwind.

Installing it

There’s no plugin system involved – it’s a straight file drop into your project root:

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

Or, if you’ve already got a CLAUDE.md, append to it:

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

It ships an equivalent Cursor project rule too, so the same guidelines apply if you switch tools mid-project.

Next Sunday: Prism Scanner, a security scanner built specifically for auditing agent skills, plugins, and MCP servers before you install them.