What problem does it solve? Autonomous agent workflows often complete tasks without any quality assessment, so regressions and weak strategies go unnoticed. This Skill provides deterministic, rule-based self-evaluation and group-relative comparison of solution strategies without requiring an external judge AI. ## Core Features & Use Cases - Self-Rewarding Evaluation: Scores completed tasks across four weighted dimensions (accuracy, completeness, efficiency, satisfaction) on a 1-5 scale with letter grades. - GRPO Candidate Comparison: Generates multiple candidate strategies or team compositions, scores them with deterministic rules (exit code, error count, speed, brevity, side effects), ranks them relatively, and updates strategy weights. - Improvement Loop: Compares scores against the last 50 evaluations, identifies weak dimensions, and generates actionable improvement suggestions when scores drop below 3.0. - Use Case: After finishing a build task, evaluate the result, compare five candidate approaches via GRPO, persist updated weights to grpo-history.json, and get a recommendation for the best team composition for the security domain next time. ## Quick Start Ask the agent to evaluate the just-completed task across all four quality dimensions and suggest improvements based on recent evaluation trends.
