What problem does it solve? AI assistant configurations accumulate rules, hooks, and skills that compensate for older model weaknesses, and nothing ever removes them. This Skill audits that scaffolding at a model upgrade and proposes deletions backed by evidence, keeping the configuration lean and justified. ## Core Features & Use Cases - Compounds vs. compensates classification: Sorts each rule into knowledge that compounds (repo-specific facts) versus scaffolding that compensates for model weaknesses and should decay. - Evidence-based audit procedure: Re-runs the failure each rule guards against, queries OTEL for 30-day invocation counts of skills, agents, and memory entries, and reads candidates cheapest-to-reverse first. - Structured removal proposals: Emits one block per candidate with location, original rationale, evidence it is no longer needed, reversal command, and risk if wrong — it proposes, never applies. - Use Case: After upgrading to a new model, run this pass to discover that a retry-on-classifier-denial hook and tool-routing prose in CLAUDE.md are no longer needed, and produce a reviewable list of deletions with rollback commands. ## Quick Start Run the scaffolding delete pass to audit my hooks, CLAUDE.md rules, skills, and agents and propose what the new model no longer needs.
