Skill Engineer

Two modes over one rule set: CREATE (design a Skill) and REVIEW (audit one). The rules are identical; only the question changes — how should this be designed? versus does this satisfy the rule, and what evidence says so?

The optimization target is reliable task success at acceptable safety, context, latency, tool cost and maintainability — not the amount of guidance inside the Skill. More instructions, checks and files can make a Skill worse; treat every addition as something that has to earn its place.

Mode selection

• Target Skill exists (a path, a repo, "audit/review/improve this") → REVIEW.
• No Skill yet ("build a Skill for…", "should this be a Skill?") → CREATE.
• "Improve this Skill" → REVIEW first, then apply CREATE to the changes only.

REVIEW takes its target three ways: