# Karpathy Skills

> Behavioral guidelines to improve Claude Code and Cursor behavior, inspired by Andrej Karpathy's observations on LLM coding pitfalls.

## Docs

- [Use Karpathy Skills in Cursor: rules and skills setup](https://docs.karpathy.phronesisagent.com/guides/cursor.md): Apply the four coding guidelines in Cursor using a committed project rule or a personal skill file, keeping behavior consistent across Claude Code and Cursor.
- [Customize Karpathy Skills with project-specific rules](https://docs.karpathy.phronesisagent.com/guides/customization.md): Merge Karpathy Skills with project-specific instructions in your CLAUDE.md to tailor AI coding behavior for your tech stack, conventions, and team standards.
- [Karpathy Skills examples: before-and-after code patterns](https://docs.karpathy.phronesisagent.com/guides/examples.md): Real code examples showing what LLMs do wrong and how the four Karpathy guidelines fix it—covering assumptions, overengineering, diff bloat, and vague tasks.
- [Install Karpathy Skills: Claude Code and Cursor setup](https://docs.karpathy.phronesisagent.com/installation.md): Install the Karpathy Skills plugin via the Claude Code marketplace in two commands, or add the CLAUDE.md file directly to any project.
- [What is Karpathy Skills? Four LLM coding guidelines](https://docs.karpathy.phronesisagent.com/introduction.md): Karpathy Skills installs four behavioral guidelines that stop LLMs from making wrong assumptions, overengineering, and editing code they shouldn't touch.
- [Goal-Driven Execution: give LLMs verifiable success criteria](https://docs.karpathy.phronesisagent.com/principles/goal-driven-execution.md): Goal-Driven Execution transforms vague tasks into verifiable goals with explicit success criteria, letting Claude loop independently until the task is done.
- [Simplicity First: write minimum viable code for LLMs](https://docs.karpathy.phronesisagent.com/principles/simplicity-first.md): Simplicity First enforces minimum viable code. No speculative abstractions, no unrequested features, and no error handling for impossible scenarios.
- [Surgical Changes: edit only what the request requires](https://docs.karpathy.phronesisagent.com/principles/surgical-changes.md): Surgical Changes restricts every code edit to the exact lines required by the request. No drive-by refactoring, reformatting, or unrelated cleanup is permitted.
- [Think Before Coding: surface LLM assumptions first](https://docs.karpathy.phronesisagent.com/principles/think-before-coding.md): Think Before Coding stops Claude from silently assuming ambiguous requirements. It surfaces tradeoffs, asks clarifying questions, and pushes back when needed.

## Optional

- [GitHub](https://github.com/forrestchang/andrej-karpathy-skills)