What problem does it solve? Correctness gates like make verify prove code works but say nothing about whether a UI is well-designed, a CLI is pleasant, or copy lands. This Skill closes that gap by iterating on subjective quality with measurable scores instead of vibes. ## Core Features & Use Cases - Adversarial loop: A generator builds or improves the artifact while a fresh, independent evaluator exercises the running result and scores it against a weighted rubric, so nothing grades its own work. - Weighted rubric with stop conditions: Define 3-5 weighted dimensions in Phase 0, then iterate until the score clears a threshold (default 7.0), plateaus (< 0.3 improvement for 2 rounds), or hits the iteration cap (default 8). - Use Case: Refine a new CLI's ergonomics by having the evaluator actually run commands, flag concrete defects like "--help doesn't list the sync subcommand", and iterate until the weighted score passes 7.0. ## Quick Start Ask the agent to run the GAN loop on your target, for example: refine the settings screen UI with a threshold of 7.5 and a maximum of 6 iterations.
