One email each morning — yesterday's AI, sortedGet it in your inbox

METAL LAB

multica-ai/andrej-karpathy-skills

205,079+589 today

A single instruction file that stops AI coding assistants from making silent bad guesses

Andrej Karpathy pointed out that LLM coding assistants tend to guess silently, overbuild simple solutions, and touch code they shouldn't. This repository packages a single CLAUDE.md file with four rules that directly counter those habits when used with Claude Code or Cursor. It is installed either as a Claude Code plugin or by appending the file to a project's existing CLAUDE.md.

What it does

  1. Karpathy observed that coding LLMs make silent assumptions, overcomplicate code with unnecessary abstractions, and sometimes edit unrelated code or comments as side effects.
  2. The repo turns these observations into four written principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution.
  3. Goal-Driven Execution reframes vague tasks like 'fix the bug' into verifiable ones like 'write a test that reproduces it, then make it pass,' so the model can loop on its own until it meets a clear success check.
  4. Installation is either a Claude Code plugin (marketplace add + install commands) or copying the CLAUDE.md file into a project; a matching Cursor project rule is also included.
  5. Success is measured by outcomes like smaller diffs, fewer overcomplicated rewrites, and clarifying questions appearing before mistakes rather than after.

Why it matters

Anyone using AI coding agents daily runs into the exact failure modes Karpathy described — silent wrong guesses, bloated code, and unwanted edits — so a drop-in guideline file offers a low-effort way to reduce review overhead and costly mistakes. The repo's 205,079 stars suggest wide adoption interest among developers who rely on Claude Code or Cursor for real work.

Terms in this repo

  • CLAUDE.md · Claude Code가 프로젝트를 다룰 때 참고하는 지침 텍스트 파일
  • Claude Code · Anthropic의 터미널 기반 AI 코딩 에이전트 도구
  • Cursor · AI 코드 자동완성/에이전트 기능이 내장된 코드 에디터
  • Goal-Driven Execution · 모호한 지시 대신 검증 가능한 성공 기준을 주고 AI가 스스로 반복 시도하게 만드는 방식

Repository description (English)

A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.

Open on GitHub

Trending repos

All repos →

Latest from METAL LAB