Pixels to Product
Design from context, not from fashionable defaults. A visual implementation is unfinished until expected and actual can be compared.
Route
- Inspect product context: tokens, shared components, typography, icons, assets, routes, representative screens, and technical constraints.
- If a visual target exists, preserve it and use it. If this is greenfield with no target, invoke
prototype-interfaceto generate exactly three genuinely different image-based directions, then wait for the user to select one. - Invoke
design-languageto translate the selected direction into semantic tokens, component rules, content hierarchy, states, and accessibility requirements. - Invoke
prototype-interfaceto produce the interactive prototype and verification matrix. - If production implementation is requested, invoke
build-vertical-slicefor the smallest real user flow. - Capture expected, actual, and diff at the same viewport, browser, scale, theme, data, fonts, and motion state.
- Invoke
review-changewith visual, interaction, accessibility, and code evidence.
Required states
Choose the applicable states and record omissions:
- narrow and wide viewports;
- empty, loading, error, partial, and success;
- hover, focus, active, selected, and disabled;
- long content, localization, font fallback, and zoom;
- reduced motion and keyboard-only navigation.
Quality bar
- Use the product’s real visual language for an existing product.
- Use semantic tokens instead of scattered values.
- Preserve readable hierarchy, consistent rhythm, and intentional density.
- Treat clipping, overflow, low contrast, invisible focus, tiny targets, and inconsistent components as failures.
- Avoid decorative gradients, glass panels, excessive cards, neon dark themes, and generic “AI startup” styling unless the selected target explicitly requires them.
- Do not claim pixel accuracy without same-context comparison artifacts.
Evidence packet
Return:
- selected target and why it fits the product;
- token and component artifacts;
- implemented routes and states;
- expected, actual, and diff image paths;
- accessibility checks and interaction tests;
- exact build and test commands with outcomes;
- remaining visual differences and untested environments.