Use when a frontend build needs to be iteratively judged on how it actually looks rendered, not just whether the code compiles or lints clean — a single page/component, or many built in parallel, each needing genuine visual polish rather than "plausible AI output." Covers the screenshot self-critique loop: a small headless-browser harness renders the real page at multiple viewports and scroll depths, the agent reads its own screenshots with vision and critiques them like a hostile design director, fixes what it finds, then adds one deliberate complexity upgrade so iteration converges on distinctive rather than bland-safe. Also covers scaling the same loop to N parallel bu…
