Andrej Karpathy skills
Karpathy-inspired Claude Code guidelines from multica-ai/andrej-karpathy-skills, think first, keep changes simple and surgical, and drive work with testable goals.
andrej-karpathy-skills packages coding-agent behavior inspired by Andrej Karpathy’s observations on LLM coding pitfalls. It is not a single “trick” skill for one stack: the core is a shared CLAUDE.md-style set of rules (also shipped as a Claude Code plugin and a Cursor rule) that push the model toward calmer, smaller, more honest edits.
What problem it targets
In short, models can assume without checking, over-complicate solutions, and change code or comments they do not fully understand. The repo turns that into four operating principles (tables and detail are in the upstream README):
| Principle | Intent |
|---|---|
| Think before coding | State assumptions, surface ambiguity, show tradeoffs, ask instead of guessing. |
| Simplicity first | Minimum code for the request; no speculative abstractions or “flexibility” you were not asked for. |
| Surgical changes | Touch only what the task needs; do not “improve” neighbors or reformat for taste. |
| Goal-driven execution | Prefer testable success criteria and verify (e.g. tests-first for bugs and behavior). |
A line often quoted in the project: LLMs are good at looping until they meet specific goals—so give clear success criteria instead of only imperative steps.
Install
Claude Code plugin:
/plugin marketplace add multica-ai/andrej-karpathy-skills/plugin install andrej-karpathy-skills@karpathy-skillsPer-project CLAUDE.md: download or append the text from the raw CLAUDE.md in the repository so the same guidance loads with your other project rules.
Cursor: the project ships .cursor/rules/karpathy-guidelines.mdc; see CURSOR.md in that repo for reuse in other projects.
Tradeoff
The upstream docs note these guidelines bias toward caution on non-trivial work—for tiny edits you can relax the full rigor so you are not over-processed for one-line fixes.