GitHub Trending: multica-ai/andrej-karpathy-skills — Andrej's rules for agents

November 5, 20262 min readGitHub Trending · AI Agents · Claude Code1 view

multica-ai/andrej-karpathy-skills is the highest-starred repo on this week's trending page: 229,663 stars, 3,808 of them this week — and it contains exactly one interesting file. It's a single CLAUDE.md, distilled from Andrej Karpathy's public observations about how LLMs write code badly. No framework, no service, no lock-in.

The four principles

Think before coding — don't pick an interpretation silently and run with it. State assumptions explicitly, present multiple readings when ambiguous, push back when a simpler approach exists.
Simplicity first — minimum code that solves the problem. No features beyond what was asked, no abstractions for single-use code, no 'flexibility' that wasn't requested. If 200 lines could be 50, rewrite it.
Surgical changes — touch only what you must. Don't 'improve' adjacent code or comments you don't understand; clean up only orphans YOUR change created.
Goal-driven execution — transform imperative tasks into verifiable success criteria ('write tests for invalid inputs, then make them pass'), then loop until they pass.

Why it's trending this week

Karpathy is one of the most respected voices in AI, and this file is his most-distributed thought in weeks. It works because it's brutally obvious in hindsight — every agent failure you've experienced maps to one of the four principles — and because installing it takes ten seconds.

Who it's for

Anyone using Claude Code, Cursor or another coding agent daily — this is the cheapest behavior fix available.
Teams whose agents keep over-engineering and rewriting half the file.
People who want a code review's worth of agent behavior without a framework.

What to know

Install: drop CLAUDE.md into your project root (or append to an existing one), or install the Claude Code plugin: /plugin marketplace add forrestchang/andrej-karpathy-skills then /plugin install andrej-karpathy-skills@karpathy-skills.
Markdown, MIT licensed — read the four principles once; they fit on one screen.
It biases toward caution over speed by design — for trivial fixes, use judgment; for anything non-trivial, watch your diffs shrink and your rewrites disappear.

The best AGENTS.md files are written by people who actually ship with agents — this is the canonical example. Repository: github.com/multica-ai/andrej-karpathy-skills.

Written by

Shihab Hossain — Full-Stack Developer

I build fast, thoughtful products end-to-end — web experiences with React & Next.js, Android apps with Flutter, and the infrastructure behind them.

Work with me

All notes

The four principles

Think before coding — don't pick an interpretation silently and run with it. State assumptions explicitly, present multiple readings when ambiguous, push back when a simpler approach exists.

Simplicity first — minimum code that solves the problem. No features beyond what was asked, no abstractions for single-use code, no 'flexibility' that wasn't requested. If 200 lines could be 50, rewrite it.

Surgical changes — touch only what you must. Don't 'improve' adjacent code or comments you don't understand; clean up only orphans YOUR change created.

Goal-driven execution — transform imperative tasks into verifiable success criteria ('write tests for invalid inputs, then make them pass'), then loop until they pass.

What to know

Install: drop CLAUDE.md into your project root (or append to an existing one), or install the Claude Code plugin: /plugin marketplace add forrestchang/andrej-karpathy-skills then /plugin install andrej-karpathy-skills@karpathy-skills.

Markdown, MIT licensed — read the four principles once; they fit on one screen.

It biases toward caution over speed by design — for trivial fixes, use judgment; for anything non-trivial, watch your diffs shrink and your rewrites disappear.

The best AGENTS.md files are written by people who actually ship with agents — this is the canonical example. Repository: github.com/multica-ai/andrej-karpathy-skills.