The CLAUDE.md File That Ate GitHub: 56K Stars, Agent Guardrails as Config

A single markdown file. 56,000 stars. multica-ai/andrej-karpathy-skills gained 42,000 stars in one week, making it the fastest-growing repo I've tracked this year that isn't an actual software product.

The file is derived from Karpathy's observations about LLM coding pitfalls. Drop it in your project root as CLAUDE.md and Claude Code automatically picks up the constraints. It targets five specific failure modes: making wrong assumptions without checking, over-complicating code, silently modifying unrelated files, not managing confusion actively, and being sycophantic to bad approaches.

I use a version of this in every project. My own CLAUDE.md is checked into this repo. The behavioral difference is significant. Without it, Claude Code tends toward "helpful assistant" mode, adding unnecessary abstractions, cleaning up code you didn't ask it to touch, saying "Of course!" before implementing something you both know is wrong. With the file, it surfaces assumptions explicitly, pushes back when warranted, and stays in scope.

What's happening here is bigger than one file. Addy Osmani's agent-skills repo hit 17,200 stars (+5,600 this week) with 20 production-grade engineering skills organized by development phase. Multica itself surged to 15,800 stars (+10,000 this week) as an open-source managed agents platform. ByteRover CLI 2.0 at 4,500 stars provides persistent structured memory for coding agents with a context tree architecture, scoring 96.1% on the LoCoMo benchmark.

The pattern is clear: agent configuration is becoming its own category. Not plugins. Not extensions. Config files that shape agent behavior the way .editorconfig shapes your IDE. The difference is that a misconfigured editor gives you wrong tab stops. A misconfigured agent gives you wrong architecture.

What builders should do: create a CLAUDE.md (or equivalent rules file for Cursor, Windsurf, etc.) for every project you touch. Don't copy-paste generic ones. Write constraints specific to your codebase. Document the assumptions your agent shouldn't make, the files it shouldn't touch, the patterns it should follow. Ten minutes of config saves hours of cleanup.