audit-repo

Audit the agentic config for the repo I'm in. Propose only — write nothing until I approve.

1. Measure

Run claude-audit repo "$PWD" and read the output. It gives you the token cost,
byte-identical copies in other repos, push-policy conflicts, dangling /command
and mcp__*__ references, and any credentials in .claude/settings*.json.
Don't recompute what it already measured.

2. Classify every line of every instruction file

Read them in full. Assign each rule exactly one bucket, one line of reasoning:

• Project fact — build commands, service names, deploy quirks, non-obvious
  layout, which tests are slow. Keep, as terse as it can be stated.
• Observed-failure correction — exists because the model actually did the
  wrong thing here, repeatedly. Keep if it still reproduces; mark re-test if
  the rule predates the current model and name the task that would prove it.
• Generic advice — "write clean code", "handle errors", "add tests", "be
  concise". Remove; the model does this unprompted.
• Stale — names a tool, file, command, or workflow that no longer exists. Remove.
• Duplicate — same rule in another file, or restated by a skill. Keep one
  copy, at the narrowest scope that still fires.
• Capability workaround — step-by-step for something now done in one shot,
  output-format scaffolds, "think step by step". Collapse to the desired outcome.
• Contradictory — conflicts with another in-scope rule. Resolve or cut both.
• Persona/tone — "you are an expert", "be careful", "world-class". Remove.

Default is remove. An item stays only if you can name the concrete recurring
situation where the model would get it wrong without it. "Seems useful" is not
that.

3. For each project skill and subagent

Does it carry knowledge the model lacks, or is it a procedure I could ask for in
one sentence? Mostly-procedure becomes a command or is deleted. Check the
descriptions are distinct enough to fire reliably — overlapping ones mis-trigger
and are a reason to merge.

4. Propose

• Rewritten files, with before/after token counts.
• One line per removal saying why.
• Anything shared with sibling repos: say which level it belongs at instead of
  being copied.
• Anything better enforced deterministically: propose the hook, not the prose.
• Re-test items as experiments — the task to run with the rule removed, and what
  the failure would look like.

Then stop and wait.