This skill should be used when building, styling, or redesigning any kind of User Interface, including web components, pages, landing pages, dashboards, or frontend interfaces. It also applies when the user says "make this look better", "style this", "design a page for X", or asks for UI polish. It should not be used for auditing existing UI without changing it (use audit-ui), writing backend logic with no visual output, or generating design tokens without an actual UI component.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
This skill should be used when building, styling, or redesigning any kind of User Interface, including web components, pages, landing pages, dashboards, or frontend interfaces. It also applies when the user says "make this look better", "style this", "design a page for X", or asks for UI polish. It should not be used for auditing existing UI without changing it (use audit-ui), writing backend logic with no visual output, or generating design tokens without an actual UI component.
Purpose
Build distinctive, production-grade frontend interfaces that avoid generic AI aesthetics. The deliverable is real working code, not a mockup. Do not write any UI code before declaring a Design Read (step 1): most bad LLM design output comes from jumping to a default aesthetic instead of reading the brief.
Before any code, read the signals: page kind, vibe words the user used, reference URLs or screenshots, the audience, existing brand assets, and quiet constraints (public sector, regulated industry, accessibility-first, kids' products). The audience picks the aesthetic, not your taste, and quiet constraints override aesthetic preference entirely.
Then declare one line before generating: "Reading this as: <page kind> for <audience>, <register>, leaning toward <design system or aesthetic family>."
With the declaration, set three dials (1-10) and hold them for the whole task; they make downstream choices auditable instead of vibes ("why does a settings page have scroll choreography?" becomes a dial violation):
Context
Density
Motion
Richness
Admin dashboard, data-heavy tool
7-9
2-3
3-4
Internal ops tool
7-8
1-2
2-3
Customer-facing SaaS product
5-6
3-4
5-6
Marketing or landing page
3-4
5-7
6-8
Settings or configuration page
5-6
2
3
Onboarding or wizard flow
4-5
4-5
5-6
Public sector, accessibility-first
5-6
1-2
2-3
Density 1 is spacious marketing, 10 is cockpit; Motion 1 is static, 10 is choreographed scroll sequences; Richness 1 is wireframe-clean, 10 is textured and layered.
If the read genuinely diverges into two directions, ask exactly one clarifying question ("closer to Linear-clean or Awwwards-experimental?"). If you can infer confidently, declare and proceed; do not ask reflexively.
2. Pick the register
Every UI task belongs to one of two registers. They have opposite failure modes, so the rules differ; read the matching reference file before building:
Dashboards, admin, SaaS apps, dev tools, internal tools
Earned familiarity: would a fluent Linear/Stripe/Raycast user trust it?
references/product.md
Mixed projects (marketing site + app) pick the register per surface, not per project.
3. Choose the foundation
If the brief reads as an established ecosystem, install and use the official package; do not recreate its CSS by hand, and do not import its tokens only to override most of them:
Brief reads as
Use
Microsoft / enterprise SaaS
@fluentui/react-components
Material / Google-flavored
@material/web + Material 3 tokens
IBM-style B2B analytics
@carbon/react
Shopify app surfaces
Polaris (required for Shopify admin)
Atlassian-style product
@atlaskit/*
GitHub-style devtool
@primer/css (Brand variant for marketing)
UK / US public sector
govuk-frontend / uswds
Modern SaaS you own
shadcn/ui or Radix Themes + Tailwind; never ship shadcn in default state
One design system per project. Aesthetics (glassmorphism, bento, brutalism, editorial) are not systems; build those with native CSS or Tailwind and say so honestly. Before importing any third-party package, check it exists in package.json; output the install command first if missing.
4. Build with the universal rules
These apply in both registers.
Laws of UX (build directives). The named cognitive and perceptual principles behind the rules below; apply them as you build. Full catalog of all 30 with directives in references/laws-of-ux.md.
Reduce choices (Hick, Choice Overload, Miller): cap one decision point at about 5 visible options; past that, group, paginate, or progressively disclose, and mark one recommended default.
One standout (Von Restorff): exactly one primary action per view, visually distinct from secondary; when everything is emphasized, nothing is.
Reachable targets (Fitts): size frequent and primary targets generously (≥ 44px), keep destructive actions clear of the primary, and exploit edges and corners for high-traffic controls.
Fast feedback (Doherty): acknowledge every action in under 400ms; if the real result is slower, show optimistic UI or a skeleton at once, never a frozen control.
Order for memory (Serial Position): put the highest-value nav and menu items first and last, the least important in the middle.
Show progress (Goal-Gradient, Zeigarnik): every multi-step flow shows steps done and remaining, makes the finish feel near, and resumes after interruption.
Group by perception (Gestalt: Proximity, Common Region, Similarity, Uniform Connectedness): signal relatedness with spacing, a shared container, and consistent styling before reaching for labels; style same-function elements identically.
Match expectation (Jakob, Mental Model): standard components behave the way users expect from the platform and category leaders; spend novelty on the brand layer, not the controls.
Forgiving input (Postel): accept input liberally (trim, normalize, accept formats), emit output strictly; never reject on a technicality you could normalize.
Absorb complexity (Tesler): irreducible complexity lives in the system as defaults and inference, not pushed onto the user as configuration.
Consistency locks. Pick once, lock page-wide, audit before shipping:
One accent color. A warm-grey site does not get a blue CTA in section 7.
One corner-radius system (all-sharp, all-soft 12-16px, or a documented per-element rule). Round buttons in a square layout is broken design.
One theme. Sections never invert from dark to light mid-page; the user must not feel they changed websites mid-scroll.
One icon family (Phosphor, HugeIcons, Radix, or Tabler; standardize strokeWidth). Never hand-roll SVG icon paths.
At most 3 font families (display + body + optional mono). Pair on a contrast axis (serif + sans, geometric + humanist), never two similar-but-different sans.
Accessibility baseline, without being asked:
Contrast: body text ≥ 4.5:1, large text and UI components ≥ 3:1. This includes placeholder text and button labels; muted-gray-on-tinted-white body copy is the single biggest reason AI designs feel hard to read. Gray text on a colored background looks washed out: use a darker shade of the background's own hue instead.
All interactive elements keyboard-reachable; visible styled focus states (never outline: none without replacement); labels associated with inputs; meaningful alt text; ARIA only where native semantics fall short.
prefers-reduced-motion alternatives for every animation. Non-negotiable.
For full WCAG 2.1 AA verification (contrast, screen readers, ARIA live regions, forms, testing tools, anti-patterns), read references/accessibility.md; interaction.md owns the keyboard, focus, and touch-target mechanics.
Serif discipline. A serif typeface is permitted only when one of two gates is met: the brief explicitly names a serif, or the aesthetic genuinely calls for one (editorial, luxury, publication, manuscript, heritage, vintage) and you can state in one sentence why this serif fits this brand. Everything else (creative agency, design studio, modern brand, premium consumer, portfolio, lifestyle) defaults to a sans display face. "It feels creative, premium, or editorial" is not a gate; reaching for a serif on a generic creative brief is a top AI design tell, because the model's reflex equates "creative" with "serif."
Typography numbers:
Body line length 65-75ch. Hierarchy through scale and weight with ≥ 1.25 ratio between steps.
Display ceiling: clamp() max ≤ 6rem. Above that the page is shouting. Letter-spacing floor ≥ -0.04em.
text-wrap: balance on h1-h3, text-wrap: pretty on long prose. No all-caps body copy.
Test headings at every breakpoint; overflowing text is a shipped bug, the viewport is part of the design.
Motion engineering:
Default to transform and opacity. The hard rule is never animate layout-driving properties (width, height, top, left, margins) casually; expand and reflow with FLIP-style transforms or grid-template-rows instead. Atmospheric properties (blur, backdrop-filter, clip-path, masks, shadows) are legitimate when bounded to small isolated areas and verified smooth on target devices.
window.addEventListener("scroll") is banned (per-frame jank): use useScroll(), ScrollTrigger, IntersectionObserver (unobserve after firing once), or CSS scroll-driven animations.
Never track continuous input (mouse, scroll progress) in useState; it re-renders the tree per frame and collapses on mobile. Use motion values (useMotionValue / useTransform).
Every animation must be justifiable in one sentence as hierarchy, storytelling, feedback, or state transition. "It looked cool" means delete it.
Reveal animations must enhance an already-visible default. Gating visibility on a class-triggered transition ships blank sections in hidden tabs and headless renderers, where transitions never fire.
Duration by purpose, not taste: 100-150ms instant feedback (press, toggle), 200-300ms state changes (menu, tooltip, hover), 300-500ms layout changes (accordion, modal, drawer), 500-800ms entrances (hero reveal only). Over 500ms for feedback feels laggy. Exits run ~75% of the matching entrance.
Ease out with exponential curves; define them as tokens (--ease-out-quart: cubic-bezier(0.25, 1, 0.5, 1), --ease-out-expo: cubic-bezier(0.16, 1, 0.3, 1)), never CSS default ease. No bounce or elastic in product UI.
Stagger is for siblings (cards in a grid, list items), never whole sections. animation-delay: calc(var(--i) * 50ms) with capped total: 10 items at 50ms is the ceiling; more items means smaller per-item delay, not a longer wave.
will-change only on elements about to animate (:hover, an .animating class), never preemptively across the page; it costs memory per layer.
States and feedback. Implement the full cycle, not just the success state: skeleton loaders matching final layout shape (not generic spinners), composed empty states that show how to populate, inline error states, and tactile :active feedback (scale-[0.98]). Two CTA hard rules, both pre-flight fails: a CTA label must fit on one line at desktop width (shorten the label or widen the button, never constrain CTA max-width), and one CTA intent gets one label across the whole page (nav, hero, footer), so "Get in touch" and "Let's talk" co-existing is a defect, not stylistic variety.
Interactive components. When the task includes forms, buttons, modals, dropdowns, tabs, or any keyboard-driven component, read references/interaction.md before building: it covers the eight-state matrix, :focus-visible rings, validate-on-blur, the native <dialog>/inert/Popover APIs that replace z-index and focus-trap hacks, undo-over-confirm, roving tabindex, and 44px touch targets.
Content and copy:
Re-read every visible string before shipping. Rewrite anything grammatically broken, referent-unclear, or LLM-poetic. Boring copy beats AI-cute copy.
Realistic data: no "John Doe", no "Acme", no 99.99% fake-perfect numbers, no invented spec precision the brand never claimed.
Button labels are verb + object ("Save changes", not "OK"). Link text has standalone meaning ("View pricing", not "Click here").
No marketing buzzwords (seamless, empower, supercharge, next-generation, elevate). Name what the product literally does.
No aphoristic cadence as the page voice: the "serious statement, then punchy short negation" rhythm is an LLM signature. Three or more sections landing on a rebuttal-shaped sentence means rewrite. Specific beats aphoristic.
Headings in sentence case, not Title Case On Every Word. No exclamation marks in success messages; confident, not loud.
Code quality: semantic HTML elements; CSS custom properties for every design token; semantic z-index scale (dropdown, sticky, modal, toast, tooltip), never z-[9999]; min-h-[100dvh] never h-screen (iOS address bar); CSS Grid over flexbox percentage math; mobile collapse declared explicitly per multi-column section; fonts via next/font or self-hosted @font-face with font-display: swap, max 2 families 2-3 weights loaded.
5. Pre-flight check and visual verification
Before delivering, verify mechanically:
Design Read was declared and the output matches it, including the three dials
Consistency locks hold: one accent, one radius system, one theme, one icon family
Contrast passes for body, placeholders, buttons, and forms
Every animation motivated, inside its duration band, and reduced-motion wrapped
Interactive elements have focus-visible rings and 44px touch targets; full state matrix on forms and overlays
Laws of UX directives applied: choices capped with a default, one standout action, feedback under 400ms, progress shown on multi-step flows, perceptual grouping
Any serif passes a gate (named in brief, or aesthetic fits with a stated reason); no default serif on a generic creative brief
No CTA label wraps at desktop; no two CTAs share one intent
Loading, empty, and error states present (product UI)
Copy self-audit done; no banned patterns from the register reference
No text overflow at 375px, 768px, 1280px
Register reference checklist passed
When browser tooling is available (screenshot, Playwright, dev server), render the result and look at it in both light and dark modes before declaring done. A UI verified only by reading its code is unverified.
Absolute Bans (both registers)
Ban
Why
Side-stripe borders (border-left > 1px as colored accent on cards/alerts)
Never intentional; use full borders, tints, or nothing
Gradient text (background-clip: text)
Decorative, never meaningful; emphasize with weight or size
Glassmorphism as default visual language
Rare and purposeful, or not at all
Hero-metric template (big number + small label + gradient accent)
Reads amateurish, not whimsical; if real assets are unavailable, ship no illustration
repeating-linear-gradient stripe backgrounds
Pure generated decoration with no brand meaning
"X theater" / "not just X, it's Y" copy constructions
Meta-criticism phrasing is instant slop; name the specific thing
The AI Slop Test
If someone could look at the interface and say "AI made that" without doubt, it failed. Check at two altitudes:
First-order reflex: could someone guess the theme and palette from the product category alone? (Cookware brief = beige + brass; fintech = navy + gold; AI tool = dark + purple glow.) If yes, rework until the category does not predict the design.
Second-order reflex: could someone guess it from category plus anti-reference ("AI tool that's not SaaS-dark, so editorial-typographic")? The first dodge is also a reflex. Rework until neither answer is obvious.
Redesign Mode
When modifying an existing UI, classify first: preserve (modernize without breaking the brand) or overhaul (new visual language, content and IA preserved). If ambiguous, ask once.
Audit before touching: extract existing brand tokens (colors, type, radii), information architecture, and patterns worth preserving. Existing brand colors override the universal palette guidance; a purple brand stays purple.
Apply modernization levers in order of lift-per-risk, stopping when the brief is satisfied: typography refresh, spacing and rhythm, color recalibration, motion layer, hero recomposition, full block replacement.
Never change silently: URL slugs, primary nav labels, form field names (breaks analytics and autofill), the logo, or legal/consent copy. Do not regress existing accessibility wins.
Gotchas
The audience picks the aesthetic, not your taste. "Pick a bold direction and commit" only applies after the Design Read; conviction in the wrong direction is still wrong. A procurement-panel B2B site executed as Awwwards-experimental fails regardless of craft.
Restraint is a register, not a failure. In product UI the expressive playbook (atmospheric backgrounds, editorial type, asymmetric grids) is the anti-pattern. Recognizing which register you are in is most of the skill.
Consistency beats novelty within a page. A page with one committed accent, one radius system, and one theme reads as designed; a page with five clever ideas reads as generated.
Transitions never fire in headless renderers. Content gated behind reveal-on-scroll classes ships invisible in screenshots and hidden tabs. The default state must be visible.
shadcn/ui in default state is a tell. It is a starting point you own; customize radii, colors, and typography to the project or you have shipped a template.
