HomeAboutProjectsRadarUniverseSkillsAchievementsContact
🌙
Back to the radar

andrej-karpathy-skills CLAUDE.md for AI Agents

instagramreview
Original post

TL;DR

A CLAUDE.md configuration file distilled from Andrej Karpathy's notes on LLM coding pitfalls, intended to constrain Claude Code (and other AI coding agents) to minimal, task-scoped file edits. It is hosted in the multica-ai/andrej-karpathy-skills GitHub repository. AI coding agents like Claude Code often over-reach — fixing one typo can trigger hundreds of unrelated file changes. This CLAUDE.md file provides four explicit behavioral constraints (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) to keep agents scoped to the stated task.

What it actually is

  • What: A CLAUDE.md configuration file distilled from Andrej Karpathy's notes on LLM coding pitfalls, intended to constrain Claude Code (and other AI coding agents) to minimal, task-scoped file edits. It is hosted in the multica-ai/andrej-karpathy-skills GitHub repository.
  • Who built it / maintained by: multica-ai (GitHub organization); inspired by Andrej Karpathy's notes on LLM coding habits
  • Status: unknown
  • Why it matters: AI coding agents like Claude Code often over-reach — fixing one typo can trigger hundreds of unrelated file changes. This CLAUDE.md file provides four explicit behavioral constraints (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) to keep agents scoped to the stated task.
  • How it compares to alternatives:
  • Cursor rules (.cursorrules)
  • GitHub Copilot custom instructions
  • Aider configuration files
  • OpenAI Codex system prompts
  • GitHub stars: unknown · License: unknown · Archived: unknown

Links

  • (no links found)

Research evidence

  • Review limitation: Repository identity unresolved: use extracted repository links or matching archive identity, not a guessed repository

General descriptions are reported, not independently tested. source_verified means the exact quotation was found in retrieved source text; it does not establish that an upstream claim works locally.

  • Identity: unresolved · Basis: unresolved
  • reported: The repository multica-ai/andrej-karpathy-skills has 230,000 stars on GitHub. (evidence: none)
  • reported: The CLAUDE.md file is derived from Andrej Karpathy's notes on AI coding pitfalls. (evidence: none)
  • reported: The file has four sections: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. (evidence: none)
  • reported: Adding the CLAUDE.md and starting a fresh Claude Code session causes the agent to restrict changes to only the files relevant to the stated task. (evidence: none)
  • reported: The file causes dead code to be reported but never deleted by the agent. (evidence: none)

Kickstarter guide

Pseudocode / untested proposal. Check current upstream instructions and the target project before execution.

According to the post, drop the CLAUDE.md file from the multica-ai/andrej-karpathy-skills repository into your project root, then start a fresh Claude Code session so it reads the file. The file reportedly contains four sections: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. No install command is needed — Claude Code automatically reads CLAUDE.md files at session start. The exact file contents are unverified; treat any reproduction as pseudocode until confirmed from the source repository.