learnprompt/andrej-karpathy-skills

This repository contains 14 AI agent skills, plus one routing skill, drawn from Andrej Karpathy's public posts and talks over several years. Karpathy is an AI researcher who has worked at Tesla and OpenAI and frequently shares observations about working effectively with AI tools. The project distills his recurring ideas into structured guides that an AI agent can follow directly. The README is written primarily in Chinese. Each skill has a defined trigger condition, a step-by-step procedure, and a clear outcome. The skill called agentic-engineering instructs the agent to break work into small, reviewable increments and run tests after each one, rather than making large sweeping changes at once. The llm-simulator skill has the agent roleplay multiple conflicting personas to expose assumptions in a decision, rather than producing a single agreeable answer. The autoresearch skill sets up a repeating hypothesis-to-experiment loop that can run unattended. Other skills cover topics like dependency minimalism, supply chain security, maintaining a Markdown knowledge wiki, and monthly self-assessment. A routing skill called karpathy-methodology sits above the others and selects which skills to combine based on the task type. Coding projects get agentic-engineering plus minimalism plus supply-chain-hygiene. Research tasks get autoresearch plus llm-wiki plus output-evolution. The project also defines four named workflows that chain multiple skills together: moving an idea to a shipped product, researching and publishing results, making decisions with reduced bias, and running a monthly capability review. The README includes test results showing which AI agent platforms each skill works on, covering Claude Code, Codex, Hermes, and OpenClaw. The skills use the Claude Code slash-command format and are compatible with the AgentSkills standard.