Everyone Is Sleeping on This One CLAUDE.md File!
A LinkedIn post by Shrey Shah — “it makes Claude Code work like a senior engineer instead of a vibe coder.”
Most teams use CLAUDE.md as a glorified README. A few rules, a project description, maybe a coding style note. Then they wonder why Claude still asks for hand-holding on real work.
Boris Cherny (the creator of Claude Code at Anthropic) quietly shared the workflows his team actually uses. Someone turned that thread into a CLAUDE.md you can drop into any repo.
It’s not a style guide. It’s a harness.
Most CLAUDE.md files describe the project. This one describes the loop the agent should run.
Six rules that do most of the work
- Plan Node Default — for any non-trivial task, plan first.
- (Plan/verify-driven loop — plan before code.)
- Lessons file — write down every correction. The agent reads those lessons at session start; over time the team’s knowledge compounds, mistake rate drops, you stop teaching the same lesson twice.
- Verification Before Done — nothing is marked complete without proof. Diff behavior, run tests, check logs. Ask: “Would a staff engineer approve this?” before claiming shipped.
- Demand Elegance — push for quality, not just working code.
- Branch wide work to subagents.
That’s harness engineering. You’re not writing better prompts. You’re building the loop the agent runs inside.
- Gist: lnkd.in/gdhPmXFb