Frontend

This file is a router over references/; the rules live there. The flow is ground -> brief -> build -> subtract -> verify. The bar is restraint with a visible hierarchy: one primary action, the important thing first, nothing on the screen without a job. A quiet, correct screen that holds its hierarchy at 390 and 1440 in both themes is a finish. Atmosphere, motion, and award-grade surface work are opt-in: build them when the brief asks (a named brand, "premium", "make it stunning", a marketing page), never by default for an internal tool, dashboard, admin screen, report, or figure. An inline page shown beside a reply (a chart, table, or diagram for show_html_page, html_preview, or html_render) follows the visualize skill when it is available; either way it is exempt from the DESIGN.md gate, the Lighthouse audit, and the research lanes, and keeps the build rules, the Draft guard, and the lint.

Route: name the references, read exactly those

Before the first edit, name the references the request routes to and one reason for each, then read them. Rows stack: take every row the request matches, so a small change that also reshapes a region takes the layout row too, and any code change takes the perfection row and, on a React project, the React tooling row. A small change on an existing surface still reads only a few; only greenfield fans out. A reference you never named is one you decided to skip; one you named but never opened is a gap you still owe.

0. Ground: know the product and the bar before designing

For a new screen or flow, a redesign, or design-system work, run references/design/grounding.md first. It reads the guideline source in full, defines the ideal state for the product's personas and journeys, and writes every gap into a ledger. The brief and the build close rows of that ledger. A small change reads the ledger when DESIGN.md has one.

1. Brief: decide what the screen is for before drawing it

Write this for every screen you build or change, into DESIGN.md section 5 beside the screen's components (the Layout: line goes under section 4, as references/design/stylegallery.md says), or into the change description when the project has no DESIGN.md. The build follows it; /visual-qa checks the rendered screen against it.

• Primary action. Exactly one. It gets the prominent style; its peers differ by style, never by size.
• Information priorities 1-3. In reading order. Priority 1 is the first thing the eye lands on.
• Left out on purpose. What a generated version of this screen would add and this one will not: the decorative badge, the stat tile nobody acts on, the subtitle under every heading, the callout that repeats the banner.
• Required content. Every input, datum, label, and piece of copy the request, the reference, or the existing screen requires, listed by name. Subtraction is checked against this list; nothing on it may disappear or merge.
• Data shape -> component (data screens): several things to compare -> table; a ranking -> ordered list; one verdict -> conclusion banner first, evidence under it; many items -> filter + results grid; one object -> detail view with its primary action. Choose from this before reaching for cards.
• Layout line (any layout or structure change, an existing DESIGN.md included): one Layout: line in the format and with the unreachable-host form given in references/design/stylegallery.md, written before the code.
• Before captures (changing an existing UI): fix the cases now, states x widths, and capture before through /visual-qa Before/after mode BEFORE the first edit. A baseline reconstructed later is not a baseline.

2. Build

Take the branch in references/design/README.md Phase 0 first (a small change on a surface with no DESIGN.md follows its small-change rule), and on a React project its Phase 0.5 tooling check: DESIGN.md (or, under the small-change rule, the styled sibling) exists before components do, and every colour, size, and spacing value traces to a token in it. A new system is built in this order: colour roles in both themes, then type and spacing, then components, which start from shadcn/ui anatomy (or the project's own kit) restyled with the tokens, never invented shapes. Then build to these rules; /visual-qa verifies them afterwards and should find nothing new.

• Hit targets. Touch layouts: at least 44x44 pt (44 CSS px) per interactive element, 8-12 pt between neighbours. Pointer-only: at least 24x24 CSS px (WCAG 2.5.8), about 10 px beyond the glyph. Grow the hit area with padding or a pseudo-element, never the glyph; adjacent bar buttons share continuous hit areas.
• Hierarchy. The brief's primary action is the one prominent control; the most important content sits top-leading in reading order; each screen's title says what it is; neither a destructive action nor Cancel is the Enter default.
• States. A control ships with hover, pressed, focus-visible, and disabled states beside its default. It keeps one box in every state: the label, icon, and spinner swap inside slots that exist in all states (a min-width sized to the longest label, a spinner overlaying the hidden label), so width, height, and padding never change when the state does, and an arrow or spinner never adds padding when it appears. A data view ships empty (offering the next action), loading (a placeholder at once, a spinner only past about a second), error (cause beside the object), and offline.
• Keyboard. Every action reachable in reading order with a visible focus ring that no sticky bar or overlay covers; Escape closes the topmost overlay; Enter triggers the non-destructive default; standard shortcuts keep their meaning; icon-only controls carry an accessible name and a tooltip.
• Motion. Only for a cause (an interaction, a state change, an arrival), on transform, opacity, filter; interruptible, never gating input; movement becomes a cross-fade under prefers-reduced-motion. A hover that changes nothing or motion on a non-interactive element is slop: leave it out.
• Icons. One SVG set (Lucide, Heroicons, Radix, Phosphor); never an emoji.
• Browser surfaces. Text selection, the caret, visible scrollbars, the focus ring, link underline offset, and tabular numerals are themed from the palette in both themes; none ships as the browser default. references/perfection/craft-floor.md has the property for each.
• Type and spacing floor. Each heading step (display to h3, and the smallest heading to body) differs by at least 1.25x, and body-family roles that sit closer also differ by weight, colour or case; the space above a section heading is at least 1.5x the space below it; the gap between groups is at least 2x the gap inside one; body text runs 65-75 characters per line on desktop; display type stays at or under 6rem and tracking at or above -0.04em unless the chosen style route sets its own. Check computed values, not the source.
• Concrete reference = contract. Match its pixels, copy, component structure, and responsive intent unless the user accepts a deviation.
• Never weaken UX to buy points. No dropping animations, hiding content, or simplifying interactions for a Lighthouse score or a deadline.

Draft guard: tells kept out of the first draft

Some tells regenerate by default, so they are kept out of the first draft rather than left for QA to find.

• No coloured bar on one side of a box, rounded or square, and no accent outline on a rounded card, decorative or state. A card, callout, warning, cover block, stat tile, testimonial, alert, or list row never carries a coloured stripe on one side: border-left: 4px solid <accent> or border-l-4 border-indigo-500, a coloured ::before/::after bar inside it, an inset box-shadow stripe, or a hard-stop gradient band or gradient border-image along one edge. A rounded card never wears a 2px-or-wider accent outline either (border: 2px solid var(--primary), ring-2 ring-indigo-500, an inset ring). Both read as generated whether they decorate the box or mark it selected, focused, or active, and an accent underline under a heading, tab, or link counts too: use a neutral 1px rule, or set the text itself in the accent ink. Give the card its identity another way: a tonal surface (one ink at a low alpha, or the next step of the surface ramp), a heading set in the accent ink, a leading glyph, or more space between groups. Encode state with one ink at many alphas (hover/selected/active washes), a check glyph for selection, and tonal layering for focus. The coloured edges that stay allowed: keyboard focus-visible rings, validation states on form inputs, neutral 1px hairlines and dividers between rows or sections, table rules, and a blockquote's neutral rule. This covers pre-existing instances on any surface you touch.
• Lint before visual QA. Run bun $SKILL_DIR/scripts/perfection/check-draft-guard.ts <changed files or dirs> on every stylesheet, component, and HTML file you wrote, including a report or figure you will print, screenshot, or render to an image. It reads CSS/SCSS/Less, HTML, JSX/TSX, Vue, Svelte, Astro, Tailwind classes (including @apply and cva/cn strings), inline styles, and styled-components, and flags the stripes above plus promotional filler copy, emoji used as icons, hype badges, and hard-coded vanity numbers in stat tiles. Exit 1 means a tell is in the draft: remove it, do not argue with the finding.
• Detect after the build renders. Once the page renders, run the detect SDK on it at 1440 and 390 with reflow on (references/perfection/detect.md): detectPage({ kind: "url", url }, { reflow: true }) from code, or bun $SKILL_DIR/scripts/detect/cli.ts --reflow <url-or-html-files>. It reads the rendered page (contrast on the painted background, type scale, heading order, overflow, decoration and copy tells). Fix every error and warning before /visual-qa; record a decision for each advisory.
• No decorative badge without a state behind it. A pill, chip, or tag appears only when it carries a value that can change: a status, a count, a category the user filters by. "New", "Pro", or "AI-powered" on a static heading is decoration.
• No stat tile without a decision behind it. A number earns a tile when someone acts on it; four tiles above a table that already holds the numbers are filler.
• No invented proof. Testimonials, quotes, customer logos, ratings, user counts, and prices come from the project or stay out. A section that needs proof the project does not have is cut, or ships as an empty slot labelled for the owner to fill; a "sample" testimonial is still a fabricated one.
• No prop decoration. A device tilted for effect, a floating reassurance card, decorative coordinates, sparkles, and ornaments that say nothing about the product are the generated version of "premium". A product preview shows a real state of the product, with its text readable at the rendered size.
• No repeated formula. A treatment that carries one heading (a serif line with an italic accent phrase, a gradient word, a headline-subtitle pair) becomes a template when every section repeats it; so does one image reused across sections. Use each once, where it means something.
• No subtitle under every heading. A heading explains itself; a one-line description under each one is the generated rhythm. Keep a subtitle only where the heading alone would be misread.
• One fact, one place. A conclusion banner and a callout that restates it, a summary card and a table row with the same figure, a label repeated as its placeholder: keep one.
• No dated or borrowed surface. Glass and blur panels (they read as a clone of a popular open-source chat app), neumorphism, mesh-gradient heroes, heavy drop shadows, and 2019-era SaaS illustration date a draft on sight. Surfaces come from DESIGN.md section 7; a material ships only when the brief or a concrete reference supplies it.
• No generated scaffold. No kicker label above a heading, no icon tile stacked over a heading in a grid of same-size cards, no hero metric (a giant number with a small label and supporting stats), no card inside a card, no neutral gray text on a coloured surface, no monospace as a "technical" costume, and no zero-offset coloured glow standing in for a shadow. references/perfection/craft-floor.md gives the replacement for each, and says when the brief or its chosen route brings one back.
• No promotional copy. Interface copy states what a thing is or does. "Seamlessly", "effortlessly", "smart", "powerful", "at a glance", "supercharge", and their Korean counterparts ("한눈에", "손쉽게", "스마트한", "효율적으로", "간편하게") are marketing, not interface: rewrite as the plain fact or delete.

3. Subtract

After the draft renders and the lint exits 0, walk every element on the screen and ask what the user could not do or know without it. No answer: delete it, whatever the reference or the previous version showed. Then open the brief's required-content list and confirm every item is still on the screen by name, with its own control or field. Subtraction that merges two required inputs into one, drops a datum, or shortens required copy is a regression, not a cleanup. Record what you removed under "Left out on purpose". Then count the controls: a secondary action most visits do not need moves into a menu or to the object it acts on, and groups separate by whitespace, not by boxes and borders. Fewer controls with room around them is the target.

4. Verify

Done is the /visual-qa gate, not your own glance. It owns the visual checks; this skill keeps build rules. Hand it the brief (B7 checks the primary action and priorities), the Layout: line (A12), the required-content list (Pass A), and, for a changed surface, the before set so it runs Before/after mode and the pair shows what improved. Its review loop is bounded (its Step 4 sets the budget): defects earn another round, and taste findings are recorded and stop it. A concrete reference runs reference-fidelity mode. Fix passes are scoped: after the first build, each fix names one change and the regions it owns (a token fix may own every region) and leaves everything else untouched, because a whole-page "make it better" pass breaks what was already right. Pass design-system decisions, evidence, and unresolved debt into /review-work for significant work.

Every UI or design-system deliverable states its WCAG 2.2 AA result. Probe the built surface in the real browser and put a table in the deliverable (PR body, change description, or reply): one row per item below, PASS / FAIL / N/A, with its evidence. A FAIL is fixed, or recorded as accepted debt in DESIGN.md section 8 with the user's sign-off.

• Contrast (1.4.3, 1.4.11): text 4.5:1 (large 3:1), icons and control edges 3:1, both themes, placeholder text included.
• Focus (2.4.7, 2.4.11): Tab through every control and screenshot each ring; none missing, none entirely hidden by a sticky bar or overlay (the AA line; the target is none covered at all).
• Target size (2.5.8): every pointer target at least 24x24 CSS px, or an undersized one whose 24 px circle, centred on it, intersects no other target and no other undersized target's circle. Exempt: inline links in a sentence, targets the browser draws unstyled, a target with a same-function equivalent that passes, and a size the information requires.
• Keyboard (2.1.1, 2.1.2): every action works from the keyboard, and focus never gets trapped.
• Reflow and text spacing (1.4.10, 1.4.12): at 320 CSS px no content scrolls sideways, except content that needs two dimensions (a data table, a map, a diagram); with line-height 1.5, paragraph spacing 2em, letter spacing 0.12em, and word spacing 0.16em applied, nothing clips or overlaps.
• Dragging, authentication, help, entry (2.5.7, 3.3.8, 3.2.6, 3.3.7): a single-pointer alternative for every drag, no cognitive test to sign in unless an alternative exists, paste and password managers work, or the test is object or personal-content recognition, help in the same place on every page, nothing asked twice in one flow. N/A only when the surface has none of these.
• State-stable controls (build rule above): read each control's getBoundingClientRect() in every state (idle, hover, loading, success, disabled); width and height match across states.
• System and surfaces (build rules above): colour roles exist in DESIGN.md before any component; each component names its shadcn/ui or project-kit source; a grep of the shipped styles finds no backdrop-filter, neumorphic double shadows, or mesh-gradient backgrounds unless the brief or a concrete reference supplies them.

On the grounded path, close the gap ledger row by row: each fixed gap has its before/after pair and the check that proves it.

The reference library (references/)

design/ holds one architecture file (design-system-architecture.md: the DESIGN.md structure, 8 sections plus a greenfield-only Research Log and the grounding gap ledger), 12 taste skills (Layer A: how to execute), 70 brand design systems (Layer B: what it looks like), and the project-original routers in the table above. _INDEX.md catalogs all 83 with mood-to-brand mappings; read it when routing is not obvious. Most non-trivial work loads one Layer A + one Layer B. README.md carries the DESIGN.md gate, the style routing flow, stacking rules, and Phase Final.

Layer A (pick AT MOST ONE style skill; they encode opposing philosophies):

Layer B: when the user names a brand or site ("Linear-style", "like Stripe's landing", "Aside-style browser agent"), load references/design/<brand>.md as the token source of truth: aside apple stripe linear.app notion vercel claude figma airbnb nike tesla spotify raycast revolut and ~56 more in _INDEX.md. Extract the tokens and apply them to the project's own content; never copy logos or trademarked imagery. A missing brand falls back to a Layer A mood match or to the nexu-io/open-design library (https://github.com/nexu-io/open-design).

perfection/: README.md (the seven tenets: real-browser audits only, 100 in every category, fix at the architecture, never weaken UX for points, design-system compliance, the audit report format) and react-perf-tooling.md (Playwright + playwright-lighthouse + react-scan/lite injection, per-route render budgets). Audit CLI, production build only, mobile and desktop presets, 3-5 runs, take the median:

uv run $SKILL_DIR/scripts/perfection/lighthouse-audit.py https://localhost:3000

ui-ux-db/: README.md documents the search CLI and the master-plus-overrides persistence pattern. The CLI finds data/ next to itself, so run it from the user's project directory: --persist writes design-system/ into the current directory, and the skill directory can sit inside a signed app bundle that must never change. Keep -B for the same reason:

python3 -B $SKILL_DIR/references/ui-ux-db/scripts/search.py "<query>" --design-system -p "Project"   # design-system generation
python3 -B $SKILL_DIR/references/ui-ux-db/scripts/search.py "<query>" --domain <domain>             # product style typography color landing chart ux react web icons google-fonts
python3 -B $SKILL_DIR/references/ui-ux-db/scripts/search.py "<query>" --stack <stack>               # html-tailwind react nextjs vue svelte astro swiftui react-native flutter shadcn jetpack-compose

designpowers/: README.md routes the pinned Owl-Listener/designpowers corpus (personas, accessibility and cognitive constraints, critique, debt, handoff, synthetic user testing, motion, role prompts) into DESIGN.md first, then into the implementation contract. It replaces nothing: not this skill, /visual-qa, /ulw-plan, /ulw-execute, or /review-work.

When to load something else instead

Activation

Use for any frontend, web UI, UX, visual, design, styling, layout, animation, performance, accessibility, or SEO work, and for every HTML/CSS/SVG artifact that ends up as a PDF, print page, report, figure, diagram, slide, or image: building, redesigning, auditing, or generating mockups. Not for backend, CLI, or pure-logic tasks with no visual surface.