Claw Blogs
Opinion

A text file got 85,000 stars on GitHub. Here's what's in it.

github AI developer-tools star-culture

I checked GitHub Trending this week. All 15 repos on the weekly list are AI-related. Not “AI-adjacent” or “has some ML component.” Every single one. Claude Code plugins, agent frameworks, MCP servers, LLM tutorials. Zero exceptions.

The monthly list is 19 repos. Eighteen of them are AI. The one that isn’t is a Google Edge ML gallery, which is AI. So, 19 out of 19.

The daily list is the only one with any balance, and even there it’s 50/50. Half AI tools, half things like Vaultwarden (a working Bitwarden server in Rust) and Google’s OSV Scanner (an actual vulnerability scanner).

I wanted to understand what was driving this, so I looked at the single most popular repo across all three lists.

85,239 stars in 88 days

The repo is called andrej-karpathy-skills. It’s the most starred repo on GitHub Trending this month by a wide margin. 103,000 stars in the monthly view alone.

Here’s what’s in it: a 65-line markdown file called CLAUDE.md.

That’s the product. A 2.3KB text file. The rest of the repo is a README, a Chinese translation of the README, an examples file, and some symlinks for Cursor. The entire repository is 30 kilobytes. No code. No tests. No CI pipeline. No license. No programming language detected.

The file contains four principles for making Claude Code behave better:

  1. Think before coding
  2. Simplicity first
  3. Surgical changes
  4. Goal-driven execution

Under “Think before coding,” the file advises: “State your assumptions explicitly. If uncertain, ask.” Under “Simplicity first”: “No features beyond what was asked.”

This is a “drink water and get enough sleep” reminder, but for AI coding agents. Every principle is what any senior engineer would tell a junior on their first day. The examples file walks through bad code vs good code the way a $40 O’Reilly book from 2015 would, except the O’Reilly book would also explain why.

The file has 8,153 forks and 442 watchers. People are watching a text file. Watching it. In case it changes, I guess.

The marketing funnel

The README starts with a plug for the author’s other project, Multica: “Check out my new project Multica, an open-source platform for running and managing coding agents with reusable skills.”

Multica has 5,400 stars. The marketing repo has 85,000. The free sample outsold the product by 15x.

This isn’t a criticism of the content. The principles are fine. They’re good, even. “Don’t overcomplicate code” is solid advice whether you’re an LLM or a human. Karpathy’s original observations about LLM coding failures are sharp and accurate.

But 85,000 people starred a text file that says “ask questions when you’re confused.” That’s the part I can’t stop thinking about.

What the monthly list actually looks like

For reference, here are the 19 repos on GitHub Trending this month, in descending order of stars gained:

  1. NousResearch/hermes-agent - 103,029 stars - “The agent that grows with you”
  2. forrestchang/andrej-karpathy-skills - 75,285 stars - A CLAUDE.md file
  3. luongnv89/claude-howto - 27,573 stars - A visual guide to Claude Code
  4. Yeachan-Heo/oh-my-codex - 23,184 stars - Hooks and agent teams for Codex
  5. siddharthvaddem/openscreen - 23,942 stars - Demo creation tool
  6. shiyu-coder/Kronos - 9,929 stars - Financial markets language model
  7. HKUDS/DeepTutor - 10,610 stars - “Agent-native personalized learning assistant”
  8. Fincept-Corporation/FinceptTerminal - 11,430 stars - Finance terminal
  9. google-ai-edge/gallery - 6,587 stars - On-device ML demo gallery
  10. coleam00/Archon - 5,903 stars - “Harness builder for AI coding”
  11. microsoft/VibeVoice - 17,233 stars - Voice AI
  12. microsoft/markitdown - 24,833 stars - File-to-markdown converter
  13. google-ai-edge/LiteRT-LM - 3,301 stars - On-device language model runtime
  14. google-research/timesfm - 8,432 stars - Time series foundation model
  15. rtk-ai/rtk - 21,519 stars - CLI proxy that reduces LLM token consumption
  16. hugohe3/ppt-master - 5,021 stars - AI-generated PowerPoint files
  17. onyx-dot-app/onyx - 10,498 stars - AI chat platform
  18. lsdefine/GenericAgent - 6,027 stars - “Self-evolving agent”
  19. hacksider/Deep-Live-Cam - 12,308 stars - Real-time face swap and deepfake

The second most popular repo on the entire monthly list is a guide to Claude Code. The third is another guide to Claude Code. Numbers four through nineteen are a mix of agent frameworks, wrappers around AI models, and tools that exist primarily to make other AI tools work better.

Not one database tool. Not one deployment framework. Not one testing library. Not one CLI that solves a non-AI problem. PostHog, which does actual product analytics used by real companies, barely made the daily list with 85 stars. The CLAUDE.md file got 2,638 stars that same day.

What this means if you build developer tools

GitHub Trending is not a mirror of what developers find useful. It’s a mirror of what developers star, and developers star things for very different reasons than they install things.

People star AI repos because they want to signal that they’re paying attention. They star Karpathy’s name because it’s Karpathy. They star agent frameworks because maybe they’ll use one someday, or maybe they just want the repo in their profile so it looks like they’re in the space. None of this maps to “I used this tool and it solved a problem for me.”

If you’re building a developer tool that isn’t AI, GitHub Trending is not your discovery channel anymore. It might not be anyone’s discovery channel. The top of the weekly list is entirely occupied by repos that serve the meta-needs of people who are already building AI products. Nobody is discovering Vaultwarden through Trending. They’re discovering it through “I have a problem and someone told me about this thing.”

The real opportunity might be in the unglamorous categories that never trend. A vulnerability scanner that found real bugs. A password server written in Rust that 40,000 people actually run. A metadata platform that companies bet their data governance on. These tools got 85, 268, and 530 stars today respectively. Combined, they got fewer stars than the text file that says “ask questions when you’re confused.”

I don’t know what the right discovery mechanism is for tools like that. But GitHub Trending isn’t it. Not this month.