andrej-karpathy-skills

Overview

A set of Claude Code guidelines inspired by Andrej Karpathy's observations on common LLM coding pitfalls. The project distills four core principles into a single CLAUDE.md file: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. These principles directly address issues like silent wrong assumptions, overengineered abstractions, unnecessary side-effect edits, and vague success criteria. It can be installed as a Claude Code plugin or appended to any existing CLAUDE.md file.

Key Features

• Think Before Coding forces the LLM to surface assumptions, present multiple interpretations, push back when simpler approaches exist, and ask for clarification instead of guessing silently.
• Simplicity First combats overengineering by prohibiting speculative features, single-use abstractions, unrequested configurability, and error handling for impossible scenarios.
• Surgical Changes ensures diffs stay minimal by limiting edits to only what the user requested, matching existing code style, and cleaning up only orphans created by the current change.
• Goal-Driven Execution transforms imperative instructions into declarative goals with verification loops, leveraging the LLM's strength at iterating until specific success criteria are met.
• Available as both a Claude Code plugin (via marketplace) and a standalone CLAUDE.md file for per-project use, with additional support for Cursor via committed project rules.

Who is this for?

• Developers who use Claude Code or similar LLM-based coding assistants and want to reduce noisy diffs, overcomplication, and silent assumption errors.
• Teams that want a lightweight, zero-config guideline file to improve AI-assisted code quality across projects without changing their workflow.
• Engineers who have experienced LLM-generated code that is bloated, makes unasked-for changes, or breaks adjacent functionality, and want a structured set of constraints to prevent those patterns.