Score UI design quality and fix AI-generated tells

Reviews a UI file or directory against a concrete design rubric on the premise that UI reads as "AI-generated" not because components are ugly but because the parts don't agree with each other - mixed corner radii, three accent colors, pure-black text, no hierarchy, missing states, robotic copy. It scores seven categories against a 100-point total, starting each at full marks and deducting for evidence-cited violations: Coherence (20 pts, the number-one "AI-generated" tell) penalizes mixed corner radii, two-or-more accent colors, emoji used as UI icons (flagged specifically for injecting uncontrolled hues versus a single line-icon set in currentColor), mixed shadow languages, mixed icon families, and inconsistent control heights; Color discipline (16 pts) penalizes pure black text instead of a refined near-black (~#2A2A2A), hardcoded hex where a semantic token exists, a normal/default state shown in a status color instead of neutral grey, status color applied to most or every row instead of just the minority needing attention, decorative per-item hues, color-only status with no icon or text, and sub-WCAG-AA contrast; Hierarchy & typography (16 pts) penalizes a number-to-unit size ratio far from roughly 2:1, uniform size/weight with no clear primary element, arbitrary font sizes with no scale, and wrong line-height; Layout & spacing (12 pts) penalizes bare-background content not in cards, off-grid spacing (7/13/19px instead of an 8px scale), a group's outer gap not exceeding its inner gap, and repeated identical section types in a row; States (12 pts) penalizes missing empty/loading/error states on data surfaces and unhelpful empty/error copy; UX writing (12 pts) penalizes buttons that don't name the actual action ("Submit" instead of "Send $2,400"), blaming or system-speak error copy ("Invalid input"), and inconsistent terminology or filler words; Motion & polish (12 pts) penalizes ad-hoc fades without a consistent named motion feel, motion that delays or blocks an action, missing prefers-reduced-motion handling, and a single hard black shadow instead of a layered tinted one. Each category clamps at 0 and sums to a total, graded on fixed bands (90+ A, 80-89 B, 70-79 C, 60-69 D, under 60 F). Output is a structured score report per file - category-by-category points with line-cited evidence - followed by a fix list ordered by score gain rather than raw severity, so the highest-leverage fixes (e.g. unifying radius and collapsing to one accent for +11 points) come first, with a projected re-score. For a directory, it gives one-line scores per file plus a full breakdown for the lowest-scoring file. It never auto-edits - it measures and recommends only, applying fixes solely when asked - and is meant as a quality gate run right after generating UI, iterating the fix list until the score clears roughly 80 before the UI reaches the user; the bar is a floor to clear, not a ceiling to chase to 100.

Design Score: 72 / 100   (src/Dashboard.tsx)   C

Coherence            13/20   sharp cards (l.22) + pill buttons (l.48); 3 accent hues
Color discipline     12/16   #000 headings (l.12, 40)
States                7/12   no empty/loading state on the orders list

Use it when a React/Tailwind/HTML UI looks generic, unfinished, or "off" and the reason isn't obvious, when a design score or pre-ship check is wanted, when a user asks for UI to look professional or polished rather than AI-generated, or right after generating UI to verify it before shipping.

Input is a UI file or directory (React/Tailwind/HTML). Output is a category-by-category score out of 100 with a letter grade, line-cited evidence for each deduction, and a fix list ordered by score gain - review only, no auto-editing unless explicitly asked.

Based on StyleSeed, an open-source (MIT) design engine with a full 74-rule reference, components, brand skins, and motion patterns beyond this condensed 7-category rubric.

Developers shipping AI-generated or AI-assisted UI who want an evidence-based, scored design review that catches the specific tells of incoherent design (mixed radii, multiple accents, emoji icons, missing states) before it reaches users, prioritized by fastest path to a better score.