What problem does it solve? Single AI agents tend to praise their own mediocre output, producing generic-looking applications with broken features. This Skill separates generation from evaluation into an adversarial feedback loop, so applications are iteratively tested and improved against a strict scoring rubric until they meet a quality threshold. ## Core Features & Use Cases - Three-Agent Architecture: A Planner expands a one-line prompt into a full product spec, a Generator implements features in sprints, and an Evaluator tests the live app with Playwright. - Structured Evaluation Rubric: Scores Design Quality, Originality, Craft, and Functionality on a weighted 1-10 scale with a configurable pass threshold and iteration cap. - Configurable Harness: Environment variables control models, iteration limits, evaluation criteria, dev server command, and evaluation modes (playwright, screenshot, code-only). - Use Case: Give a one-line brief like "Build a project management app with Kanban boards and dark mode" and let the loop run 5-15 iterations until the live application passes the rubric. ## Quick Start Ask the agent to run the GAN-style harness to build a recipe sharing platform, starting with a planning pass and then iterating generation and Playwright-based evaluation until the weighted score reaches 7.0.
