What problem does it solve? Improving prompt or instruction prose is usually ad hoc: edits pile up without measurement, regressions go unnoticed, and there is no way to restore a known-good version. This Skill turns instruction improvement into a bounded, auditable experiment loop where every candidate is measured against a frozen evaluation suite and only kept when it genuinely improves. ## Core Features & Use Cases - Bounded atom selection: Classifies targets as deterministic code (routed to ordinary tests), bounded instruction atoms (eligible), or multi-atom orchestration (refused), so experimentation never becomes an unbounded rewrite. - Frozen evaluation oracle: Freezes the suite identity and digest before mutation, measures a baseline, and decides KEEP or REVERT per candidate using declared metrics, non-regression rules, and budget gates. - Auditable ratchet with restoration: Persists an audit row for every candidate, restores exact last-KEEP bytes after REVERT, enforces time/iteration/token/cost/diff budgets, stops on plateau, and requires human review before integration. - Use Case: You want to improve one paragraph of a review prompt. The loop freezes your eval suite, tries one bounded candidate at a time, keeps only measured improvements, reverts regressions automatically, and hands the cumulative diff to a human for final approval. ## Quick Start Use the autoresearch-loop skill to iteratively improve this instruction paragraph against my frozen evaluation suite with a KEEP/REVERT ratchet and a five-iteration budget.
