Karpathy's 227k-star skills library: 109 PRs are adding features to a file that says 'don't add features'

There is a GitHub repository called andrej-karpathy-skills with 227k stars. What it does is simple: it distills Andrej Karpathy's observations about the mistakes LLMs make when writing code into a CLAUDE.md file to guide Claude Code's behavior.

Karpathy previously tweeted about several problems: models make wrong assumptions on your behalf and then run with them, don't manage their own confusion, don't seek clarification, don't surface inconsistencies, don't present trade-offs, and don't push back when they should; they like to overcomplicate code and APIs, inflate abstractions, not clean up dead code, and write 1,000 lines for something that could be done in 100; they sometimes also modify or delete comments and code they don't fully understand, even when unrelated to the task.

This repository distills these problems into four principles and writes them into a file:

1. Think Before Coding: Don't assume, don't hide confusion, lay out the trade-offs. If unsure, ask; if ambiguous, list multiple interpretations; if there is a simpler approach, push back; if confused, stop.
2. Simplicity First: Solve the problem with the minimum amount of code; don't write anything speculative. Don't add unrequested features, don't write abstractions for one-time use, don't provide unrequested "flexibility", don't write error handling for impossible scenarios. If 200 lines can become 50, rewrite it.
3. Surgical Changes: Touch only what must be touched, clean up only the mess you made. Don't "improve" adjacent code, comments, or formatting, don't refactor what isn't broken, match the existing style. If you find unrelated dead code, mention it; don't delete it.
4. Goal-Driven Execution: Define success criteria and loop until verification passes. Turn "add validation" into "write tests for invalid input, then make them pass"; turn "fix a bug" into "write a test that reproduces it, then make it pass."

But the most interesting part isn't the file itself. After Niels Rogge tweeted about this repository, Daniel Lenton replied underneath: "109 open PRs on a file that says no features beyond what was asked". A guide teaching people not to over-engineer is itself facing 109 feature-adding PR requests. This is probably the most fitting dark humor of the LLM era.

The README plainly states "No features beyond what was asked", and yet people are rushing to add features to it. It's like writing a book about minimalism and every reader wants to stuff a chapter into it.

Repository address: https://github.com/multica-ai/andrej-karpathy-skills