[EUREKA] Karpathy's LLM Coding Rules in your claude code now

You know that feeling when an AI assistant writes code that works—but leaves behind unnecessary abstractions, dead code, and a structure only it understands? You accept the PR because it passes tests, then spend hours cleaning up. LLMs don't ask for clarification, they make assumptions, and they overbuild. Karpathy identified six specific failure modes: wrong assumptions, no confusion management, overcomplicating code, bloated abstractions, and dead code that never gets cleaned.

Think of it like a project manager for your AI coding assistant. You drop a 65-line CLAUDE.md file into your project root. Every time you ask Claude Code to do something, it reads these four rules first: understand before changing, keep it simple, make minimal changes, and define what 'done' looks like. The file doesn't add features—it shapes behavior. Like giving a junior dev a checklist: 'Before you code, explain what you're about to do and why.'

• Think Before Coding — saves you from AI making changes it doesn't understand: Claude reads your codebase, identifies impact zones, and explains its plan before writing a single line
• Simplicity First — cuts through AI's tendency to over-engineer: no unnecessary abstractions, no 'frameworks' for simple problems, dead code gets deleted immediately
• Surgical Changes — prevents cascading messes: Claude touches the minimum files needed, understands dependencies first, and tests incrementally instead of rewriting your entire codebase
• Goal-Driven Execution — stops micromanagement hell: you define success criteria, Claude loops until it meets them, no step-by-step babysitting required
• Zero Setup — works instantly: copy one file to your project root, or install as a Claude Code plugin, no configuration needed
• MIT License — free to modify and redistribute: adapt the principles to your team's workflow, no licensing friction

Who It Is For

If you're using Claude Code and tired of cleaning up over-engineered AI commits, this is for you. Perfect for solo developers or small teams who want guardrails without heavyweight processes. Not useful if you're on Cursor, Copilot, or other AI tools—this only works with Claude Code. Also not a fit if you've already trained yourself to give better prompts; the file automates what you might alread...

Worth Exploring?

Yes, if you use Claude Code and feel the pain of over-engineered AI suggestions. The principles come directly from Karpathy's real workflow shift from 20% to 80% agent coding—he's not theorizing, he's sharing what actually works. The 9.4k stars and HN discussion prove real developers are finding value. But understand what you're getting: a behavior-shaping file, not a new tool. If Claude Code's built-in Plan mode already works for you, this adds minimal overhead. Start with the examples in EXAMPLES.md to see if the style matches how you think.