What problem does it solve? Optimizing prompts, code, configs, or creative text by hand is slow and unmeasurable — you tweak, eyeball the result, and never know if you actually improved. This Skill turns a fuzzy optimization goal into an approved research brief with a concrete metric, baseline, and target, then runs an automated generate-score-verify-evolve loop until the target is met, the round budget is spent, or progress plateaus. ## Core Features & Use Cases - Metric-driven research loop: Define a metric, evaluator backend (programmatic command, LLM judge panel, or human scoring), baseline, and target, then evolve candidates across rounds with genetic mutation, recombination, and wildcard exploration. - Adversarial verification: Every round winner is attacked by a refuter checking constraint violations, metric gaming, harness bugs, and prompt injection before its score stands. - Auditable, resumable runs: Every round persists candidates, scores.yaml, verify.yaml, and round logs to disk, producing a final dossier, leaderboard, and best-artifact copy; runs resume via --resume. - Use Case: Optimize a prompt until accuracy reaches 0.90, get p95 latency under 200ms, or evolve taglines until a judge panel scores 8.5 — all with a full provenance trail. ## Quick Start Ask the agent to run autoresearch with the goal of optimizing your prompt until its accuracy on the eval set reaches 0.90.
