What problem does it solve? Rapid UI prototyping often produces generic, AI-slop-looking interfaces with no objective quality bar, and self-evaluation by the same agent that built the app is unreliable. This Skill runs an adversarial loop where a generator builds a live app and a separate evaluator drives it in a real browser, scores it against a weighted rubric, and feeds concrete fixes back until a quality threshold is crossed. ## Core Features & Use Cases - Three-phase adversarial loop: Plan derives scope from real sources (a Decision Register, mined specs, or an explicitly gated exploratory proposal), Generate builds and commits per iteration, and Evaluate drives the live app and scores design, originality, craft, and functionality on a weighted rubric with a 7.0 pass threshold. - Honest evaluation modes: detects the available browser-automation driver at runtime (Playwright MCP, Chrome DevTools MCP, desktop automation) and records the mode actually achieved — live-driver, screenshot, or code-only — never silently downgrading. - AI-slop craft checklist: a concrete checklist of generic gradients, unthemed component libraries, stock heroes, and missing interaction states, with citations used in evaluator feedback. - Use Case: You want to explore what a dashboard for an existing brownfield app could look like. The loop pulls requirements from mined specs, builds a live prototype, drives it in a browser, scores it, and iterates up to 5 rounds until the weighted score reaches 7.0. ## Quick Start Ask the assistant to run the gan-harness loop to prototype and iteratively improve a UI for the current project until the design rubric score passes 7.0.
