AI/TLDR

Forrest Chang · 2026-01-27 · notable

andrej-karpathy-skills — a single CLAUDE.md for better Claude Code behavior

A single drop-in CLAUDE.md file encoding four rules for steering Claude Code — think first, keep it simple, make surgical changes, verify against explicit success criteria. 16.4k stars, +4.9k this week.

andrej-karpathy-skills GitHub repository social card

A drop-in CLAUDE.md that teaches Claude Code to stop overreaching and start verifying.

Key specs

LicenseMIT
GitHub stars16.4k
Stars this week+4.9k

What is it?

A single markdown file you either install as a Claude Code plugin or curl into a project as CLAUDE.md. Forrest Chang distilled it from watching LLMs burn context on the same coding failure modes — unexamined assumptions, over-refactoring, and declaring success without evidence.

How does it work?

The file encodes four principles that Claude Code ingests at every session: 'Think Before Coding' forces explicit reasoning and clarifying questions, 'Simplicity First' bans speculative code and premature abstractions, 'Surgical Changes' restricts edits to what the user actually asked for, and 'Goal-Driven Execution' turns imperative tasks into declarative success criteria that have to be verified before the agent claims done.

Why does it matter?

Coding agents routinely over-refactor, invent unused helpers, and report success without running tests. This file is the terse, opinionated corrective that a lot of Claude Code power users were writing by hand. It crossed 16k stars in a few weeks — adoption is happening because it works and you can apply it in one command.

Who is it for?

Anyone using Claude Code daily.

Try it

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

Sources

Tags

  • claude-code
  • claude-md
  • coding-agents
  • prompt-engineering
  • developer-workflow
  • open-source

← All releases · Learn AI