Design Score
tells you
what's wrong.
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
- For a full prose audit with fixes → use
- For logic/config with no visual artifact — scoring is meaningless
Step 0 — Resolve the effective rule set
Before scoring, read
.styleseed/effective-rules.md
and
. If they are
missing or
reports drift, invoke
or
from the
project lock first. Only fall back to the source handbook when no project lock exists.
Score in authority order: core invariants first, then the exact output grammar, domain/page,
brand recipe, optional aesthetic profile, and bounded lock values. The lock cannot waive an invariant.
Unknown values are a resolver error; do not invent an exception.
The output must name the effective rule set, for example:
text
Rule set: operations-console × SaaS × dashboard × enterprise-workbench × swiss
What to score
Score the file (or each file in a directory) on eight weighted categories that
map to the design language. Total = 100.
| Category | Weight | Reads from |
|---|
| Color discipline | 16 | DESIGN-LANGUAGE §1, §18, §72 + VISUAL-CRAFT §C4 |
| Hierarchy & typography | 16 | §2, §3, §4, §16 + Font Size table + VISUAL-CRAFT §C2 |
| Layout & rhythm | 12 | §13, §14, §15, §61 + VISUAL-CRAFT §C1 |
| Cards & elevation | 10 | §7, §8, §12, §1 + VISUAL-CRAFT §C3 |
| States & a11y | 18 | §11, §70, §71, §72 + VISUAL-CRAFT §C3 |
| Motion & interaction | 6 | §24, §59 + |
| Coherence | 12 | VISUAL-CRAFT §C0 (one choice per axis) |
| Distinctiveness | 10 | Golden Rules 14–16 + VISUAL-CRAFT §CC-9b (not generic/default/template) |
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 (16) — deduct for: accidental
/
outside a profile or
grammar contract that explicitly uses hard black structurally (−4 each, cap −8); competing
decorative emphasis hues (−5);
emoji used as UI icons (−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);
the unlocked default indigo (/) used as the accent instead of
a chosen domain-fit color (−4).
Distinctiveness (10) — a coherent screen can still read "AI-generated." Deduct for: the
icon-chip cliché — a generic Lucide line-icon in an identical pale-tinted rounded-square,
repeated for every feature/step (−4, §CC-9b); the StyleSeed demo layout copied verbatim
(hero+chat / 3-step / feature-grid / pricing) with no product-specific identity (−4); no focal
point — an all-even grid of same-weight, centered, evenly-spaced cards (−3); the hero shows a
stock/placeholder visual instead of this product (−3); the escape hatch as a new uniform
(§CC-9c) — ghost 01/02/03 index numbers on every section, or identical uppercase-overline +
big-number cards repeated with no variation (−2); distinctive-but-dated (§CC-9d) — full
beige/paper page base, serif body text on a product surface, dark-heavy blocks that read
"brochure" not "2026 product" (−3). Cap −10.
Hierarchy & typography (16) — deduct for: number/unit not ~2:1 (−4); font
sizes off the Font Size table /
for size (−5); everything the
same weight, no clear primary (−5); cramped or wrong line-height on body (−3);
body < 16px on a desktop/web B2B surface (tight mobile scale on a wide screen) (−4 —
but dense-data chrome is exempt: chart ticks, mono SHAs/timestamps, table metadata at
12–13px are correct; and dashboard app-chrome h1 at 22–24px is correct, not a violation
of the marketing 40–56px headline scale).
Layout & rhythm (12) — deduct for: grouping that contradicts the selected grammar or recipe (−6):
needs explicit functional groups, while
should not be
forced into cards;
needs aligned panels/rows while
needs flat step flow; arbitrary off-scale spacing (−3); same section type repeated without purpose
(−4); no discernible proximity rhythm (−3).
Cards & elevation (10) — deduct for mixed or task-inappropriate surface language. Hairlines,
flat grouping, tonal ramps, or restrained shadows are valid only when the selected grammar/profile
uses them coherently. Deduct mixed border/shadow languages (−4), visibly heavy or directionally
inconsistent shadows (−4), or missing group/surface separation where the grammar requires it (−5).
States & a11y (18) — deduct for: missing empty/loading/error state on a data
surface (−5 each, cap −10 — a static mockup or marketing landing with NO data surface is
N/A: skip these deductions, don't fail the category); contrast below 4.5:1 body / 3:1
large (−6); touch target < 44px on a touch surface (pointer-first desktop controls at
36–40px are fine) (−4); no visible focus /
(−5); icon-only control
without
(−3).
Motion & interaction (6) — deduct for: random/ad-hoc fades instead of a named
seed/keyword (−3); motion that delays content or blocks an action (−4); no
handling on custom motion (−3).
Scroll-linked/parallax/3D/animated-
gradient is SURFACE-DEPENDENT (§43): on an app/dashboard/data/form surface it's forbidden
(−5); on a
marketing/landing/brand page it's ALLOWED (the Cinematic tier) — there, do NOT
deduct for scroll-linked reveals, pinned sections, 3D hero, or animated backgrounds; only deduct
for
scroll-JACKING (hijacking scroll / trapping) (−5), motion that hides content until scroll
or delays the headline/CTA (−4), or a missing
fallback (−3). Judge by
page type first, then score.
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
(−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)
Rule set: operations-console × product-ui × SaaS × dashboard × enterprise-workbench × swiss
████████████████░░░░░░ C-
Color discipline 11/16 ▓▓▓░ competing orange+blue emphasis hues (l.28-34)
Hierarchy & typography 13/16 ▓▓▓▓ number/unit 1:1 on hero (l.18)
Layout & rhythm 9/12 ▓▓▓░ two identical KPI rows (l.22-31)
Cards & elevation 8/10 ▓▓░░ mixed border + floating-shadow language (l.22)
States & a11y 11/18 ▓▓░░ no empty/loading state; focus ring missing (l.55)
Motion & interaction 4/6 ▓▓▓░ default fade, not a named seed
Coherence 6/12 ▓▓░░ sharp cards (l.22) + pill buttons (l.48); 3 accent hues (§C0)
Distinctiveness 8/10 ▓▓▓░ all-even KPI grid weakens the operational focal panel
### 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
run as a loop, not a one-off:
- Score the just-generated UI.
- If < 80, apply the "fix first" list (use to make the edits), then re-score.
- Repeat up to ~3×, or until ≥ 80.
- 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.
- Order the "fix first" list by score gain, not by severity alone — the goal
is the fastest path to a better number.
- 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. measures; and 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.