# Autoresearch Skill ## Overview Autoresearch is a stateful improvement loop with a strict evaluator contract. You define a mission (what to improve) and an evaluator (a command that measures success). The skill loops until the evaluator passes or max runtime is reached. ## The Loop ``` Mission defined → Baseline evaluation while (evaluator fails AND time remaining): 1. Generate improvement hypothesis 2. Implement change 3. Run evaluator 4. Record result in decision log 5. If better: keep change; if worse: revert 6. Repeat with new hypothesis ``` ## How to Invoke ``` Skill: autoresearch Mission: Improve Lighthouse performance score from 72 to 90+ Evaluator: npm run lighthouse Max runtime: 2 hours ``` ## Evaluator Requirements The evaluator must: - Be a runnable shell command - Return exit code 0 on success, non-zero on failure - Produce consistent, reproducible results ## Output Artifacts - `.omc/autoresearch/decision-log.md`: Every hypothesis tried + result - Final state of improved code - Summary of changes made