# andrej-karpathy-skills

> A CLAUDE.md template fixing common LLM coding pitfalls using four principles.

This repository provides a single CLAUDE.md configuration file designed to eliminate frequent AI coding failure modes, including silent assumptions, code overcomplication, and unintended edits to neighboring files. Based on observations shared by Andrej Karpathy, the guidelines enforce explicit clarification, minimal code, surgical edits, and test-driven success criteria. It drops directly into any Claude Code or Cursor project with zero setup.

Website: https://github.com/forrestchang/andrej-karpathy-skills

## What it is for

- Prevent Claude Code or Cursor from making unwarranted changes to untouched code
- Stop AI coding assistants from silently guessing when given ambiguous specifications
- Force AI models to write simpler, test-verified code rather than bloated abstractions
- Standardize AI code generation practices across a developer team

## Features

- Four core principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution
- Enforces explicit questions and tradeoff analysis before implementation begins
- Prohibits speculative abstractions, unused flexibility, and drive-by refactoring
- Encourages test-first development loops with concrete verification steps
- Includes ready-to-use Cursor project rule configuration (.cursor/rules)
- Available as a single file download or installable Claude Code plugin

## Pricing

Open source

## Details

- Type: GitHub repo
- Category: [Development](https://xiry.app/topics/development.md)
- Topic: [AI Coding Prompts](https://xiry.app/topics/development/ai-coding-prompts.md)
- Open source: yes
- Platforms: Command line, other
- For: Software developers using Claude Code, Cursor, or LLM coding assistants
- Tags: claude code, cursor, llm guidelines, coding assistant, karpathy, prompt engineering

## Why it was saved

- [@dr_cintas, 3 May 2026](https://x.com/dr_cintas/status/2050910884164207063): Encodes 4 behavioral constraints to fix common LLM coding bad habits: silently making wrong assumptions, overcomplicating code, editing code that wasn't asked to change, and running without pushback or clarifying questions.
- [@dunik_7, 10 May 2026](https://x.com/dunik_7/status/2053495306235347100): The floor template referenced by the article; the full file reportedly at the end of the article.

## Related

- [Taste Skill](https://xiry.app/resources/taste-skill): Portable agent skills that enforce design quality in AI-generated frontends (GitHub repo, Free)
- [emilkowalski/skills](https://xiry.app/resources/emilkowalski-skills): AI agent skills for UI animation, design engineering, and interface polish (GitHub repo, Open source)
- [prepare-branch-context](https://xiry.app/resources/prepare-branch-context): Claude Code skill to ground the agent in the current branch diff (Prompt, Free)
- [Remotion Prompt Showcase](https://xiry.app/resources/remotion-prompt-showcase): Gallery of community prompts for generating Remotion videos with coding agents (Website, Free)
- [TAME Cursor Rules for Claude 3.7 Sonnet](https://xiry.app/resources/tame-cursor-rules-for-claude-3-7-sonnet): Cursor rules to keep Claude 3.7 Sonnet focused and non-destructive (Prompt, Free)
- [UI Skills](https://xiry.app/resources/ui-skills): Curated design-engineering skills and constraints for coding agents (Website, Free)

Updated 10 October 2026. Page: https://xiry.app/resources/andrej-karpathy-skills

Make your own library: Xiry takes your saved reels, tweets, Reddit bookmarks and Watch Later videos, makes them searchable and pulls the useful stuff out, like this library. It's invite-only for now; join the waitlist at https://xiry.app/#waitlist.