karpathy-guidelines

Overview

Karpathy Guidelines is a Claude Code skill inspired by Andrej Karpathy's observations on common LLM coding pitfalls. It provides a single CLAUDE.md file with four core principles designed to combat issues like wrong assumptions, overcomplication, bloated abstractions, and unintended side effects when AI agents write code.

Key Features

• Think Before Coding - Forces explicit reasoning by requiring the agent to state assumptions, present multiple interpretations, push back when simpler approaches exist, and stop when confused
• Simplicity First - Combats overengineering by prohibiting speculative features, unnecessary abstractions, unasked-for flexibility, and code that could be significantly shorter
• Surgical Changes - Prevents unintended side effects by limiting edits to only what the request requires, without improving adjacent code, reformatting, or deleting pre-existing dead code
• Goal-Driven Execution - Transforms imperative tasks into verifiable goals with defined success criteria and test-first approaches

Who is this for?

This skill is designed for any developer using AI coding agents who has experienced the frustration of LLMs making wrong assumptions, overcomplicating solutions, or making unauthorized side-effect changes to code. It is ideal for teams that want a lightweight, principle-based approach to improving AI code generation quality across any project.