What problem does it solve? Builds that compile but don't actually work slip through when agents stop at green tests. This Skill drives a build from "compiles" to near-finish-line craft by repeatedly running, rendering, exercising, and critiquing it against locked acceptance criteria, with explicit stop-conditions so it never thrashes forever. ## Core Features & Use Cases - Disciplined iteration loop: Repeats build → see → exercise → check → critique → rebuild, orchestrating existing tools (bash build/test, headless screenshots, Playwright, axe/Lighthouse) rather than reinventing a harness. - Dual-track ledger: Tracks objective machine facts (exit codes, test counts, console errors, a11y violations) alongside a design bar when feel is load-bearing, including a different-model critic for taste. - Guaranteed termination: Stops on PASS, PLATEAU, BUDGET (default 4 rebuild passes per surface), or BLOCKED, and hands off anything it cannot verify as a named human gate. - Use Case: After scaffolding a new web feature, invoke the loop to render key UI states across mobile and desktop, drive the core flows headless, fix the top defects each pass, and stop with a per-iteration ledger showing exactly what improved. ## Quick Start Tighten this build by looping until the acceptance criteria pass, then report the iteration ledger and which stop-condition fired.
