/polish — fix the slop, by the root

/design-check found what's wrong. This skill fixes it, the right way. The trap is decorating a screen that's actually just missing a design system. Fix causes, not symptoms.

Before you touch a pixel

• Read DESIGN.md. It's the source of truth for tokens, fonts, and rules. If there isn't one, stop and run /design-intake first, polishing without a system just produces a different flavor of default.
• Don't polish unfinished work. If the feature isn't actually working yet, finish it (/check) before making it pretty.

How to polish

1. Classify each issue (from /design-check) as one of:
   • Missing token (no defined color/size for this) → add it to DESIGN.md, then use it.
   • One-off deviation (ignored an existing token) → snap it back to the token.
   • Conceptual misalignment (fighting the intended aesthetic) → rework it toward the North Star, don't patch it. Fixing the root cause is what stops the slop from coming back next screen.
2. Run the detector, fix what it finds. node bin/detect-slop.mjs <path> lists every measurable tell with file:line and the fix. Work the list: swap the reflex font, kill the gradient text, replace the side-stripe border, drop the violet gradient, fix contrast (>=4.5:1, OKLCH), put spacing on a 4pt scale, cut buzzwords and em-dashes. Then handle the catalog tells a script can't see: card-grid monotony, imagery, hierarchy (reference/anti-slop-tells.md).
3. Check every interactive element has all its states: default, hover, focus (visible!), active, disabled, loading, error. Missing states are a top slop tell and an accessibility fail.
4. Stay restrained. Don't add animation, gradients, or glow that nobody asked for. One accent, used rarely. prefers-reduced-motion respected. The fix for "looks AI-made" is usually less, not more.
5. Prove it. Re-run node bin/detect-slop.mjs <path> until it reports 0 P1, then re-screenshot and run /design-check. The score should go up and the tells should be gone. Don't claim it's polished until the re-check confirms it.

Red flags, stop and correct

Voice

Surgical and restrained. Every change traces to a tell or a token. The goal is a screen you can't tell was AI-assisted, reached by subtraction more than addition.