A 70-Line File Just Passed 220,000 GitHub Stars
It contains no code — just four rules for how an AI agent should behave. That it's now one of the most-starred repos on GitHub tells you exactly where the bottleneck moved.
NeuroX AI · June 12, 2026
A configuration file with no executable code just passed 220,000 combined GitHub stars — roughly 91,200 on the creator's account and over 132,000 on an organization mirror. The core file is about 70 lines of Markdown.
It's a CLAUDE.md, created by Forrest Chang on January 27, 2026 — one day after Andrej Karpathy described his shift from manual coding to agent-driven development and the failure modes that came with it. Chang turned those failures into four named rules and dropped them at the root of a repo:
- Think Before Coding — state assumptions, surface ambiguity, ask instead of guess.
- Simplicity First — solve the stated problem, no speculative abstractions.
- Surgical Changes — don't touch code, comments, or formatting outside the request.
- Goal-Driven Execution — convert vague instructions into verifiable success criteria first.
It crossed 100,000 stars in three months. Not because it's clever — because each rule maps to a way agents quietly wreck production code: scope creep, invented requirements, drive-by refactors, and a "done" no one can verify.
That's the whole lesson. The model isn't the bottleneck anymore; the instructions are. A team that writes its discipline down ships agents that behave. A team that doesn't ships a demo that regresses on the second prompt.
We bake constraints like these into every agent before the first line runs.