Review and redesign an app or website UI to strip out

Unslop design

Three parts: what to fix (checklist.md, each fix with its reason), how to orchestrate it (process.md, rounds with reviewers and devs in parallel) and cases and sources (references.md: the before/afters that shaped this skill and a map of Refactoring UI). The spec card every reviewer fills is in spec-card.md. Per-platform notes (SwiftUI, Compose, web, widgets, watch, car) live in platforms.md; websites, web apps and dashboards have their own playbook in web.md. Read the one you need, not all of them.

Three scripts measure instead of guessing:

• python3 scripts/slop-scan.py <paths> — slop patterns in source (web, SwiftUI, Compose) with file, line and fix number. Standard library only.
• node scripts/web-audit.mjs <url> — the rendered page at 390/768/1280 px: horizontal scroll, a census of every size/color/radius/shadow/spacing in use, computed contrast, focus states, emoji, glows, plus a grayscale screenshot. Needs Playwright.
• python3 scripts/unslop-score.py <paths> [--audit report.json] — both as one number from 0 to 100 with where the points went, plus a README badge. Run it before and after: "score 41 → 88" is the PR's headline.

examples/dashboard-before.html and -after.html show both on one page.

Principles (non-negotiable)

1. A fix is a rule, not a tweak. Every change ships as a token or component (spacing, font size, weight, color role, radius, shadow, button), never as a literal in the view. If the token is missing, create it in the same change. Pick values by elimination within the scale (try the value and its two neighbors; the middle wins or you repeat), never pixel by pixel.
2. Hierarchy before color. The screen has to work in grayscale: spacing, size, weight and contrast do the work; color comes afterwards and only to mean something. If it's unclear in gray what matters, color won't fix it.
3. The first thing you read is the answer. One hero number per screen (the balance, the price, today's progress); everything else steps back. Emphasize by de-emphasizing the competition, not by inflating the hero. At most one dark/hero surface per screen, and none in sheets or modals.
4. Color says one thing. One brand accent for the action or the hero number; everything else is a role with meaning (money/success, up/danger, status, reward) and gray for what only navigates. Decorative color = slop. Color never works alone: shape or a word goes with it.
5. Measured contrast, not assumed. WCAG AA (4.5:1 text, 3:1 large text and icons) computed from the real hex values against the real backgrounds, light and dark. White on a green/yellow brand color almost never passes; gray text on a colored background doesn't either (pick a shade of the same hue).
6. Nothing the screen can't deliver. No invented numbers, no projection presented as data, no implied feature that doesn't exist, no empty state that lies. Design the smallest version you can actually ship.
7. One signature gesture, taken from the domain. A brand detail born from what the app shows (a price-board price with a smaller last digit, dot-matrix numerals on a running app), not a new font or effect; and it degrades gracefully where it doesn't apply. Everything else is native: don't fight the tab bar, switches or system sheets.
8. Platforms ship together. If the project has iOS and Android (or app and web), the change goes to both in the same PR, in every language.
9. Verify what the user sees. Screenshots in light and dark, largest text size, a non-default currency/language, real data (long names, empty, 2,000 items). What can't be compiled in your environment gets audited by reading, and the PR says so.

Any agent (Claude Code, Codex, Cursor, Copilot, Gemini…)

The rules are the same everywhere. Where it says "agents", use what your tool has: subagents if it has them (Claude Code, Codex…), with a git worktree each for devs. With no subagents, run the reviewers one after another, each with its own brief, and keep their outputs separate before the jury step. Where it says "publish", Claude Code publishes an artifact; other agents write an HTML file the user opens. In a chat app without a codebase, stop after the before/after proposal.

Quick mode (the default for one screen)

One screenshot, one screen or one page, or the user says "quick": no subagents, no jury, one pass you can finish in a few minutes.

1. Ground it in one line (what the screen is for, its surface mode) and measure what you can: scripts/unslop-score.py on the screen's code (with --audit for a web page). Then the passes on the render (process.md § 2): one second, squint, cover the logo, left edges, gray.
2. Pick the five fixes that change the screen most, each with its checklist.md number and one line of why. Hierarchy and spacing before color, color before polish. Say what you would not change.
3. Show the before/after: the edited code's render, or an HTML mock of the screen if there's no code, with the filled spec-card.md rows it uses. Score it on the six axes of process.md § 4 first; under 3 on any axis, revise before showing.
4. On "yes", apply them as tokens (principle 1), rerun the score and report "score X → Y" and what's still left.

Switch to the full flow below when there are several screens or videos, no design system yet, platforms to ship together, or the user asks for a full review.

Surface modes

Not every screen gets the same ambition. Name the mode before judging:

Mode	Screens	What wins	Expression budget
Operate	dashboards, settings, editors, admin	scanning, consistency, platform conventions	low: one signature at most
Persuade	landing pages, onboarding, paywalls	one message, one action, a real product shot	higher: one bold idea
Read	articles, docs, detail pages	line length, rhythm, quiet chrome	low
Experience	games, maps, camera, media	the content itself as the canvas (#59)	high, from the domain

The checklist applies to all four; the budget decides how far #47 and #58 go.

Working agreement (full flow)

• Reviewers = senior product designers and devs = senior devs, as agents (use the model the user asks for). The orchestrator doesn't build the areas: it does the shared work, integrates and audits.
• Every decision comes with an explicit, actionable recommendation.
• Before implementing, publish the before/after and wait for the user's "yes".
• Every round also reviews what earlier rounds shipped and all the videos/screenshots received, not just the latest.
• When merging is allowed once CI is green: close with a list of what to check when compiling (concrete risks) and, if there's no Mac/simulator, a self-contained prompt for a local session.
• On "is everything done?": done (with PRs) · pending on you · left out on purpose.
• After the last round, update the project's design system.

Full flow

1. Inventory: frames from the videos (process.md § 1), screens and states, and each screen located in the code. Run scripts/slop-scan.py on the UI code and, for the web, scripts/web-audit.mjs on each key page.
2. System: check (or create) the spacing, type, color and elevation scales (checklist.md § System). Without scales, reviewers judge by eye; the audit's census shows which values the page really uses.
3. Grayscale pass + parallel reviewers (3-4, one per area) against checklist.md, applying fixes one at a time.
4. Winning proposal: one HTML page with before/after, the spec card and the list of decisions; the user chooses.
5. Devs per area in isolated worktrees with a closed brief; integrate, compile, audit, PR, answer review bots, merge when green.
6. Finally: update the project's design system (doc + tokens + its own skill) with what's new.

If the project already has its own design skill

Use it for concrete tokens and components; this one brings the method and the general checklist. If a rule here clashes with a written project decision, the project wins and the proposal points it out.