AIToolScan

Karpathy Skills Overview

Karpathy Skills (multica-ai/andrej-karpathy-skills) is a viral, single-file collection of behavioral guidelines for AI coding agents, distilled from Andrej Karpathy's observations on LLM coding pitfalls. It works with Claude Code, Cursor, Codex, and other AI coding agents to stop wrong assumptions, overengineered code, and unrelated edits before they happen.

• Four Core Principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution — packaged into one CLAUDE.md file and a Claude Code plugin.
• Think Before Coding: State assumptions explicitly, present multiple interpretations, surface tradeoffs, and push back on flawed prompts instead of guessing.
• Simplicity First: Minimum code that solves the problem — no speculative features, no bloated abstractions, and no error handling for impossible scenarios.
• Surgical Changes: Touch only what you must, match existing style, and never refactor or delete unrelated code and comments as side effects.
• Goal-Driven Execution: Transform tasks into verifiable success criteria with tests-first loops so the agent can verify each step and loop until done.
• Claude Code Plugin: Install via the plugin marketplace (forrestchang/andrej-karpathy-skills) so the guidelines apply across all your projects.
• Cursor & Multi-Agent Support: Ships a .cursor/rules/karpathy-guidelines.mdc project rule for Cursor, plus a plain CLAUDE.md usable by Codex and other agents.
• Open Source: MIT licensed with 200K+ GitHub stars, one of the most popular AI coding agent guideline repositories.