Andrej Karpathy Skills is a single CLAUDE.md instruction file for AI coding assistants, with more than 224,000 stars on GitHub under the MIT license. It encodes Andrej Karpathy’s observations on how LLMs go wrong when writing code as four working principles any Claude Code or Cursor user can adopt in minutes.
Karpathy’s critique was blunt: models make wrong assumptions and run with them without checking, they overcomplicate code and bloat abstractions, and they quietly change or remove code they do not understand. This repository distills that critique into a compact set of behavioral guidelines. Drop the file into a project and the assistant is instructed to state its assumptions, prefer the simplest solution that works, touch only the lines the task requires, and define verifiable success criteria before looping toward them.
Andrej Karpathy Skills is free and open source under the MIT license. There is no paid tier, no account, and nothing to run: the deliverable is a Markdown file you copy into your own projects and freely customize.
For developers who use Claude Code or Cursor daily and are tired of bloated diffs, drive-by refactors, and silent wrong assumptions โ and for teams who want a shared, reviewable baseline for how their assistant should behave.