ui-score – Design & UI skill for AI agents

Design Score

When to Use

Use this skill when you need score a UI file's design quality 0-100 against StyleSeed's design language — per-category breakdown, the worst offenders, and a prioritized fix list. A quantified version of /ss-review.

/ss-review tells you *what's wrong*. /ss-score tells you *how good it is

overall* and *what to fix first* — a single number plus a category breakdown, so

you can track UI quality like you track test coverage.

When NOT to use

• For a quick pass/fail before committing → use /ss-lint
• For a full prose audit with fixes → use /ss-review
• For non-UI files (logic, config) — scoring is meaningless

Reference availability

DESIGN-LANGUAGE, VISUAL-CRAFT and engine/motion are not bundled here. Use the inline rubric against the actual project design system and label unsupported reference checks unverified. Gate mode does not authorize edits beyond the user’s requested scope.

What to score

Score the file (or each file in a directory) on seven weighted categories that

map to the design language. Total = 100.

| Category | Weight | Reads from |

|---|---|---|

| Color discipline | 18 | DESIGN-LANGUAGE §1, §18, §72 + VISUAL-CRAFT §C4 |

| Hierarchy & typography | 18 | §2, §3, §4, §16 + Font Size table + VISUAL-CRAFT §C2 |

| Layout & rhythm | 14 | §13, §14, §15, §61 + VISUAL-CRAFT §C1 |

| Cards & elevation | 12 | §7, §8, §12, §1 + VISUAL-CRAFT §C3 |

| States & a11y | 18 | §11, §70, §71, §72 + VISUAL-CRAFT §C3 |

| Motion & interaction | 8 | §24, §59 + engine/motion |

| Coherence | 12 | VISUAL-CRAFT §C0 (one choice per axis) |

How to score each category

For each category, start at full marks and subtract for violations you find by

reading the code. Be specific and evidence-based — cite the line.

Color discipline (18) — deduct for: any #000/text-black (−4 each, cap −8);

more than one accent hue used decoratively (−5); emoji used as UI icons (multi-color,

breaks single accent) (−5); a normal/OK/"보통" state shown in a status color instead of

neutral grey (−4); status color on most/every row (no severity hierarchy) (−4);

decorative hues (gold stars, rainbow category dots) instead of accent/grey (−3);

hardcoded hex where a semantic token exists (−2 each, cap −6); status conveyed by color

alone (−4).

Hierarchy & typography (18) — deduct for: number/unit not ~2:1 (−4); font

sizes off the Font Size table / text-[var(--…)] for size (−5); everything the

same weight, no clear primary (−5); cramped or wrong line-height on body (−3).

Layout & rhythm (14) — deduct for: content on bare background, not in cards

(−6); px-4/px-8/mx-4 instead of px-6/mx-6 (−3); same section type

repeated in a row (−4); no space-y-6 rhythm (−3).

Cards & elevation (12) — deduct for: 1px borders doing separation work that

tone+shadow should (−4); shadows over ~8% opacity / visibly heavy (−4); no

card/background tone separation (−5).

States & a11y (18) — deduct for: missing empty/loading/error state on a data

surface (−5 each, cap −10); contrast below 4.5:1 body / 3:1 large (−6); touch

target < 44px (−4); no visible focus / outline:none (−5); icon-only control

without aria-label (−3).

Motion & interaction (8) — deduct for: random/ad-hoc fades instead of a named

seed/keyword (−3); motion that delays content or blocks an action (−4); no

prefers-reduced-motion handling on custom motion (−3); scroll-linked/parallax

(forbidden, §59) (−5).

Coherence (12) — the "one choice per axis" laws (VISUAL-CRAFT §C0). Deduct for

each axis that is *mixed* rather than unified across the file: mixed radius

personalities, e.g. sharp panel + pill buttons (−5); two+ competing accent hues used

for emphasis (−4); mixed shadow languages / light directions (−3); mixed icon

families, fill modes, or stroke weights (−3); same radius on a nested element instead

of inner = outer − padding (−2); inconsistent control heights for buttons/inputs

(−2). This is the category that most predicts "looks AI-generated" — weight evidence

of system-wide consistency, not per-component prettiness.

Clamp each category at 0. Sum to a total.

Output format

## Design Score: 70 / 100   (src/app/Dashboard.tsx)

████████████████░░░░░░  C-

Color discipline      13/18   ▓▓▓░  #000 headings (l.12,40); orange+blue+green accents (l.28-34)
Hierarchy & typography 15/18  ▓▓▓▓  number/unit 1:1 on hero (l.18)
Layout & rhythm        11/14  ▓▓▓░  two identical KPI rows (l.22-31)
Cards & elevation       8/12  ▓▓░░  1px borders doing separation (l.22)
States & a11y          11/18  ▓▓░░  no empty/loading state; focus ring missing (l.55)
Motion & interaction    6/8   ▓▓▓░  default fade, not a named seed
Coherence               6/12  ▓▓░░  sharp cards (l.22) + pill buttons (l.48); 3 accent hues (§C0)

### Fix first (highest score gain)
1. Add empty + loading states to the orders list       → +7 states (§71)
2. Unify radius (pick soft 8-12px) + collapse to one accent → +9 coherence+color (§C0, §2)
3. Drop the 1px borders, use tone + ≤8% shadow         → +4 cards  (§7)

Re-score after: ~92 / 100.

Use letter bands: 90+ A · 80-89 B · 70-79 C · 60-69 D · <60 F.

Gate mode (use this as the Quality Gate before showing the user UI)

The Quality Gate (CLAUDE.md / AGENTS.md) is /ss-score run as a loop, not a one-off:

1. Score the just-generated UI.

2. If < 80, apply the "fix first" list (use /ss-review to make the edits), then re-score.

3. Repeat up to ~3×, or until ≥ 80.

4. Present the UI with the final score and a one-line "fixed: …".

The pass bar is a floor, not a ceiling — get to ≥ 80 and stop; don't chase 100. The point

is that no first-draft, obviously-incoherent UI reaches the user. Especially never ship below

80 with a rainbow status list, emoji icons, two accents, or missing states — those are the

exact tells the gate exists to catch.

Rules

• Read the file — score from real evidence (line numbers), never guess.
• Prioritize accessibility and task-blocking defects before cosmetic score gain. The number is an aesthetic heuristic, not validated quality, authorship detection or permission to ship.
• For a directory, print a one-line score per file, then the lowest-scoring file's

full breakdown.

• Don't auto-edit in plain scoring. /ss-score measures; /ss-review and /ss-motion fix.

In Gate mode (above) you do fix-and-re-score until the floor is met.

• As a *gate*, ≥ 80 is a floor before showing the user — but don't over-polish: chasing 95→100

to delay shipping is worse than shipping a clean 85.

Limitations

• Use this skill only when the task clearly matches its upstream source and local project context.
• Verify commands, generated code, dependencies, credentials, and external service behavior before applying changes.
• Do not treat examples as a substitute for environment-specific tests, security review, or user approval for destructive or costly actions.