skill-routing-audit

# Skill Routing Audit A skill can hold the right answer and still fail, because routing buries it. Happy-path "how do I X" questions route to the right reference. Boundary and negative questions ("can it even do X?", "what can't it do?", "is Y possible?") route to that same "how to do X" reference, which only lists what the subject can do, so the real "no" sits in another file or nowhere. The reader lands where the routing points and never sees it. This skill hunts that failure with boundary probes and a skill-only simulation. ## When to Use - After authoring or editing a skill, before relying on it. - To sweep every installed skill for the same class of gap. - When a skill that should cover a question gave a vague or wrong answer. Not for creating a skill from scratch (`skill-creator`) or folding a user's manual edit back into a skill. ## Method Point it at a skill's directory. Audit one named skill, or every subdirectory that has a `SKILL.md`. Fan out one read-only subagent per skill (parallel for an all-skills sweep). Each agent: 1. Read the skill's `SKILL.md`, then its reference and bundled files, to learn what it covers and what it says a user can and cannot do. 2. Derive 4 to 6 boundary/negative probes a real user would ask at the edge of the skill's scope: plausibly-impossible actions, off-limits operations, common misconceptions. Not happy-path "how do I X". Ground each in the skill's own subject. 3. Answer each probe using only this skill: start from `SKILL....