Project CLAUDE.md Audit

Audit this project's agent-instruction files. The goal is minimizing startup context — every line in a CLAUDE.md loads into every session, so the bar is high. Output a concise cut / move / keep / fix-contradiction list with file:line refs. Do not edit until approved.

Run /audit-project-harness first. Knowing what the project's hooks enforce automatically lets this pass cut the instructions they make redundant (see Step 3, "Prefer enforcement over instruction").

Be efficient: don't read whole files into the parent context. Delegate discovery and reading to an Explore subagent and work from its condensed findings + file:line refs.

Step 1 — Global baseline

Read ~/.claude/CLAUDE.md once. Treat everything in it (communication style, "read first", code-style defaults, discuss-before-implementing, backpressure/verification, the safety hooks) as ALREADY IN EFFECT. Project files must not restate or weaken any of it.

Step 2 — Locate (subagent)

Dispatch Explore to find every CLAUDE.md / AGENTS.md in the repo (root + nested) and return: the tree, plus each file's section headers and any inline content that looks long, duplicative, or derivable. Don't pull full file bodies into the parent — request excerpts only where a judgment call needs them.

If no such files exist, this skill has nothing to audit — point the user at /init-project-claude to scaffold them instead.