How Andrej‑Karpathy‑Skills Fixes Karpathy’s AI Coding Pitfalls

GitHub project andrej‑karpathy‑skills is an open‑source skill inspired by Andrej Karpathy’s X thread about the shortcomings of LLM‑generated code. Karpathy reports that 80 % of his code now comes from LLM agents, which has transformed his workflow but also exposed several fatal flaws: unchecked assumptions, over‑engineering, silent code modifications, and lack of clarification.

Four principles to tame agents

Think before coding – force the agent to state assumptions, ask when uncertain, present alternatives, and propose simpler solutions.

Simplicity first – prohibit unnecessary features, abstractions, or handling of unlikely error cases; rewrite bloated code to a concise form.

Precise edits – modify only what is required, avoid altering unrelated code or comments, and ensure every change can be traced to a user request.

Goal‑driven execution – convert imperative tasks into verifiable declarative goals with tests; let the agent iterate until the success criteria are met.

The CLAUDE.md file in the repository encodes these rules and can be applied to any project. When the agent follows them, diff noise drops, initial code is cleaner, and pull requests become tidy.

Installation

For Claude Code users, the skill can be added via the plugin marketplace:

/plugin marketplace add forrestchang/andrej-karpathy-skills
/plugin install andrej-karpathy-skills@karpathy-skills

Or download the CLAUDE.md file directly:

curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md

After adding the file, the rules are automatically applied to the project. Users can extend CLAUDE.md with project‑specific constraints such as strict TypeScript mode or mandatory API tests.

Karpathy notes that the most powerful engineers will not spend time crafting prompts but will build systems around agents, using the four principles as an operating system for reliable, high‑productivity AI‑assisted development.