claude-md-authoringlisted

# claude-md-authoring A CLAUDE.md is not free documentation. Claude Code injects it wrapped in a system-reminder that says it "may or may not be relevant," and the system prompt already spends on the order of fifty instructions before yours arrive. Every line you add competes for the model's attention with every other line, so a bloated CLAUDE.md does not make the agent more careful — it dilutes the few instructions that actually matter and quietly buries them. This skill governs how the harness writes the CLAUDE.md files it ships into consumer projects: minimal, universal, runnable, with detail deferred and certainties mechanized. **Attribution**: see `.source` in this directory. --- ## The load-bearing insight The CLAUDE.md is appended to the system prompt with a "may or may not be relevant" framing. That framing is the whole game: - It is **not** a contract the model must obey. It is context the model weighs. - Attention is finite and shared. A 400-line CLAUDE.md does not give you 400 enforced rules; it gives you 400 things competing to be the one the model attends to this turn. - So the question for every line is never "is this true?" but **"does this earn a slot against everything else here?"** Target ~60 lines. Hard ceiling well under 150 instructions. If you are past that, you are not documenting — you are hoping. --- ## What belongs in CLAUDE.md Only the **universal** and the **load-bearing**: - How to run the project: build, test, dev server, the one comma