Member-only story
The 4 Lines Every CLAUDE.md Needs
What Karpathy diagnosed, what 60,000 developers bookmarked, and why behavioral constraints beat feature checklists
In a single week in April 2026, Anthropic shipped Claude Opus 4.7, launched a new product called Claude Design, and added Routines that run even when your laptop is closed.
On the same day, OpenAI ratcheted up Codex with parallel agents that click and type on your Mac.
This is normal now. April 2026 has been called one of the most packed month for LLM releases on record. Roughly 65–70% of enterprise code is AI-written. Over 50% of companies describe their AI adoption as a “chaotic free-for-all.”
And yet the most-starred developer resource in this entire space isn’t a framework, a plugin, or a model.
It’s four sentences in a markdown file.
One GitHub repo. 60,000 stars.
No dependencies, no API, no build step.
Just a CLAUDE.md file with four behavioral guidelines derived from something Andrej Karpathy posted in January. What 60,000 developers bookmarked tells you more about the real bottleneck in AI-assisted coding than any product announcement this week.
