A CLAUDE.md File Became an Ecosystem: What Started with Karpathy Now Has 89K Stars and Dozens of Forks
On January 26, 2026, Andrej Karpathy posted a thread on X about what was happening to him after changing the way he programmed. He had shifted from 80% manual code with autocomplete to 80% AI agents in a matter of weeks, and had very specific observations about the failure patterns he kept seeing repeat.
It wasn’t a vague complaint. It was a clinical diagnosis.
The models assumed without asking and moved forward without verifying. They overcomplicated code, inflated abstractions, didn’t clean up dead code. They made “orthogonal” changes — touching things that had nothing to do with the task. And when you gave them procedural instructions, they executed them mechanically instead of understanding the actual objective.
A dev named Forrest Chang read that thread and turned each observation into a concrete instruction for a CLAUDE.md file.
What Chang Did, Exactly
The forrestchang/andrej-karpathy-skills repo is a configuration file, not a library or a framework. A file that Claude Code reads when starting each session in that directory.
Four principles coded as behavior directives:
Don’t assume. If there’s ambiguity, the model has to name it explicitly and ask. Don’t silently pick one and move on.
Don’t hide confusion. If something isn’t clear, stop. Say what’s confusing. Ask for clarification. The file states it directly: if there are multiple possible interpretations, present them — don’t silently pick one.
Show the tradeoffs. Before implementing, identify if there are simpler approaches. Minimum code that solves the problem. No features nobody asked for, no abstractions for single use.
Goal-oriented execution. The most powerful principle. Instead of procedural instructions (“do X, then Y, then Z”), give it verifiable success criteria and let the model figure out the path. “Make the test fail, then make it pass” instead of “fix the bug”.
To install it in a project:
curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md
Or to add it to an existing CLAUDE.md without overwriting:
echo "" >> CLAUDE.md && curl https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md >> CLAUDE.md
The repo also includes support for Cursor (karpathy-guidelines.mdc) and is structured as a Claude Code marketplace plugin, so you can also install it with:
/plugin marketplace add forrestchang/andrej-karpathy-skills
The Numbers That Made Noise
The repo was created on January 27, 2026. In the weeks that followed, it reached the top of GitHub’s trending, accumulating over 89,000 stars and 8,500 forks. In the week of April 22, it gained +44K new stars and returned to first place in the weekly ranking.
For a repository that essentially contains a text file, that’s a very clear signal: the problem it solves is universal.
Karpathy didn’t participate in developing the file. He didn’t publicly endorse it. Chang built the solution from Karpathy’s observations; the community validated both.
The Ecosystem That Grew Around It
Ten or more derived repos appeared in the weeks that followed. The pattern bifurcated in two directions:
Extensions of the original CLAUDE.md. The most notable is a community gist nicknamed “v2 fixclaw”, published on April 22. It adds six new rules to Chang’s four — focused on runtime, not just at coding time. It includes directives for token budgets per step, per pipeline, and per day, and a key rule: use the model only for tasks that are genuinely language tasks (classification, unstructured text extraction, summarization), not for routing decisions that an if/else could make cheaper and deterministically.
The LLM Wiki pattern. In April, Karpathy published another gist — this time about a different idea: instead of RAG (re-deriving answers from raw documents every time), use the LLM to build and maintain a structured markdown wiki. Knowledge accumulates with each source you add; each question you ask enriches the base. You don’t re-derive; you query what’s already been synthesized.
That generated another cluster of forks:
kfchou/wiki-skills— Claude Code plugin that implements the wiki pattern directlyAstro-Han/karpathy-llm-wiki— a single-skill implementation with examples and operation logslucasastorian/llmwiki— open-source implementation with UI, PDF loading, and MCP serverskyllwt/OmegaWiki— the most ambitious: 23 Claude Code skills for a complete academic research platform, from paper ingestion to paper writing and reviewer responses
Why This Matters Beyond CLAUDE.md
The andrej-karpathy-skills phenomenon is a case study in how de facto standards form in the agent ecosystem.
Karpathy named the problems with precision. Chang coded them as behavior instructions. The community validated them at scale. And now there’s a karpathy-skills tag on GitHub with dozens of repos — skills for rigorous reasoning, for coding discipline, for personal wikis, for academic research.
The file itself is a good practice in prompt engineering: you don’t tell the model what steps to follow, you define how it should behave when facing ambiguity and how it should formulate success before it starts.
That scales. That’s why it has 89K stars.
How to Use It in Your Workflow
If you’re using Claude Code in serious projects, the most direct use case is to add the file to your global CLAUDE.md (~/.claude/CLAUDE.md) so it applies to all your projects without having to install it project by project:
echo "" >> ~/.claude/CLAUDE.md && curl https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md >> ~/.claude/CLAUDE.md
If you use OpenCode, Hermes, or another compatible agent, the file works the same renamed as AGENTS.md.
And if the wiki pattern catches your attention — using it to maintain domain knowledge, internal documentation, or research bases — the most direct entry point is kfchou/wiki-skills for Claude Code or lucasastorian/llmwiki if you prefer a web UI with an MCP server.
Four principles. One text file. 89,000 stars.
Sometimes the simplest solution is the one that scales best.