Web Design

web-design

Guidance for producing web UI that looks deliberately designed, not auto-generated. The single failure mode this skill exists to prevent: shipping something a viewer can immediately tell an AI made. If it reads as the AI default — purple gradients, beige hero, three equal cards, em-dashes, Inter on slate-900 — it failed, regardless of how clean the code is.

Three ideas run through everything here:

• Taste is trained, not innate. Good design is a stack of small correct decisions, each defensible. There is no single magic move.
• Unseen details compound. The 75ms exit animation, the 1px focus ring offset, the tinted neutral instead of pure gray — individually invisible, collectively the difference between "fine" and "considered."
• Restraint beats expression by default. One accent, one type scale, one corner-radius scale, motion only where it earns its place. Reach for more only when the brief asks for it.

When to use this

Use it whenever you design or review web UI: a landing page, a dashboard, a component, a form, a marketing section. Use it before you call a screen "done." Don't treat it as always-on — it's a lens for visual/frontend work, not a tax on every code change.

Workflow

1. Read the brief (below) — infer intent, write a one-line Design Read, set the dials.
2. Apply the fundamentals — type, color, spacing, the interactive states. Pull depth from references/ as needed.
3. Avoid the anti-patterns — references/anti-patterns.md is the load-bearing don't-list.
4. Look at it in a browser — screenshot and eyeball. Do not ship on faith (see Iterate by eye).
5. Run the pre-ship checklist (below) before declaring done.

The brief read

Before writing any CSS, infer from what you were asked:

• Page kind — product UI / dashboard / landing page / form / docs / editorial. This dominates every other choice (a dashboard and an Awwwards landing page want opposite densities).
• Audience & register — internal tool / consumer / enterprise / public-sector.
• Vibe words & references — anything the brief implies about feel ("calm", "punchy", "premium").
• Hard constraints — accessibility requirements, regulated/public-sector rules, brand assets. These override aesthetic preference, always.

Then write a one-line Design Read ("Internal analytics dashboard, dense, calm, a11y-first — restrained/standard/dense") so the intent is explicit and checkable. Ask at most one clarifying question, and only when the brief is genuinely ambiguous — otherwise pick sensible defaults and proceed.

The intent dials

Three orthogonal knobs, each at three named levels. Set them from the brief; they bias every later choice. They are set conversationally — say what they are, don't write them into a file.

Default = restrained / standard / standard. This is restraint-biased on purpose: most real work is product UI and dashboards, where high expressiveness actively harms usability. Move a dial up only when the brief earns it — an expressive marketing hero is legitimate; an expressive settings page is not. Whatever the dials, a mobile single-column override and the accessibility constraints still apply.

The fundamentals (the spine)

The corroborated core. Compact here; full treatment in references/.

• Type — one modular scale, one ratio ≥1.25 (1.25 / 1.333 / 1.5); ~5 sizes cover most needs; body ≥16px; measure 45–75ch (cap 65–75ch); heading line-height 1.1–1.2, body 1.5–1.7; ≤3 font families. → references/typography.md
• Color — work in OKLCH, not HSL; 60-30-10 weight split; tinted neutrals, never pure gray; dark mode is a surface-lightness scale, not an inversion; commit to one accent. → references/color.md
• Layout & space — a 4pt spacing scale (4/8/12/16/24/32/48/64/96); tight grouping 8–12px, section separation 48–96px; deliberate hierarchy (combine size + weight + space); a semantic z-index scale (never 9999); ≥44×44px touch targets; pass the squint test. → references/layout-spacing.md
• Interaction — design all eight states (default / hover / focus / active / disabled / loading / error / success); a visible :focus-visible ring always; placeholders aren't labels; undo beats confirm; skeletons beat spinners. → references/interaction.md
• Motion — short and motivated. UI animations <300ms, exits ~75% of the enter; custom decelerating easing, never bounce/elastic, never ease-in for UI; animate only transform/opacity; respect prefers-reduced-motion. Decide whether to animate before how — and motion you commit to must actually ship: a page that reads as animated but sits static is broken, so if you can't land it cleanly, dial motion down and ship clean static. → references/motion.md
• Copy — buttons are verb+object (never OK/Submit/Yes); errors say what happened, why, and how to fix; no em-dashes; no marketing buzzwords. → references/ux-writing.md
• Accessibility — WCAG AA minimum (4.5:1 body text, 3:1 large text and UI); this is a floor, not a goal.

Avoid the AI default

The highest-value content in this skill is the don't-list in references/anti-patterns.md — a cross-corroborated catalogue of the tells that mark generated UI (AI-purple gradients, reflexive beige/cream, gradient text, eyebrow chips, numbered section markers, three equal feature cards, em-dashes, overused fonts). Read it before building and check against it before shipping. The framing: if you can tell an AI made it, it failed.

If a real design system fits, use it

If the project already uses — or the brief clearly calls for — an established design system, adopt its official package rather than hand-rolling CSS that approximates it. One system per project; don't recreate its tokens by hand, and don't import a system and then override most of it. Match the brief to the ecosystem it lives in:

(This skill is stack-agnostic and prescribes no framework or package beyond naming these systems.) If the brief is an aesthetic (glassmorphism, bento, brutalist, editorial) rather than a system, there's no official package — build it with native CSS and be honest about what's borrowed inspiration vs. official material. If nothing fits, build restrained native CSS on the fundamentals above.

Output modes

• Build / enhance — apply the rules directly to the work.
• Review / fix — emit a focused | Before | After | Why | table, one row per change. This is the default review format: concrete, scannable, each change justified.
• Full critique (on request) — a deeper pass using the compressed review lens in references/anti-patterns.md (heuristics + cognitive-load + persona lenses).

Redesigning an existing UI

A redesign is not a greenfield build — misclassifying the mode is the biggest source of bad redesign output. First decide: preserve (modernise without breaking the brand) or overhaul (a new visual language over the same content and IA). If genuinely ambiguous, ask once.

Audit before you touch anything. Record the current brand tokens (accent, type stack, radii), the information architecture (page tree, nav labels, conversion paths), and the SEO/analytics baseline — SEO and analytics migration is the #1 redesign risk. Extract the existing brand accent before applying the color rules: a brand that is already purple stays purple (the anti-AI-purple rule governs your defaults, not an established identity).

Apply the levers in priority order and stop when the brief is satisfied — each is more risk than the last: (1) typography, (2) spacing & rhythm, (3) color recalibration (unify neutrals, keep the brand accent), (4) a motion layer, (5) hero / key-section recomposition, (6) full block replacement (only for a block that's unsalvageable). Targeted evolution (1–4) is most of the value at a fraction of the risk.

Never change silently — get explicit sign-off before touching URL slugs / anchor IDs, primary nav labels, form field names or order (breaks analytics + autofill), the brand logo/wordmark, or legal / consent copy.

Iterate by eye

Looking at the rules is not the same as looking at the result. If this session has browser testing available (a browser-testing-capable profile — a QA or web-design rig with a Playwright/browser MCP — spawns with one), use it:

1. Render the page or component in the browser — at the dev server's actual bound URL. Assert the bound port from the tracked server process itself, by whatever means your OS exposes (the framework's own startup banner, when captured — e.g. vite's Local: http://…:PORT — is one way but not the only one); never assume a default port. If that port is already held by another process, the dev server binds another port or fails, so eyeballing the default would render the wrong, stale server. For a static on-disk HTML file with no dev server (a CV, a rendered report), don't navigate file:// — Playwright's browser_navigate blocks it outright — and don't hand-roll a web server per render cycle. Serve its directory over loopback with the bundled helper and open the printed URL instead: node .claude/skills/web-design/scripts/serve-static.mjs <dir>.
2. Screenshot it and actually look — squint test, hierarchy, spacing rhythm, contrast, the eight states, dark mode, a narrow (mobile) viewport. For a throwaway verification shot, take it with no filename/path — the screenshot tool auto-names it into your session's out-of-tree scratch/output dir, so the working tree stays clean. Only pass a path when you deliberately persist a shot (to attach or diff), and then make it an absolute path under the per-session scratch directory the Playwright client itself allows writes to — this is not necessarily the same as your generic session scratch/temp dir, and a path outside Playwright's own allowed roots is rejected ("… is outside allowed roots"); a bare or relative name also resolves against the browser tool's client workspace (the repo root), litters the tree, and risks an accidental commit (git status will flag it). The auto-name-into-scratch and $LOOM_SCRATCH_DIR mechanics here are @playwright/mcp-specific: only a session that mounts Playwright gets that allowed root exposed as the $LOOM_SCRATCH_DIR environment variable — persist a shot there rather than your generic session scratch/temp dir. A session on a DIFFERENT browser tool (e.g. claude-in-chrome) has no $LOOM_SCRATCH_DIR and no auto-naming — use that tool's own capture flow instead. To persist a shot as a file under Playwright, don't rely on claude-in-chrome save_to_disk — it renders the inline base64 but writes no reachable file (Claude Code issue #40141). Use Playwright page.screenshot({ path }) against the loopback page (launch with { channel: 'chrome' } to reuse system Chrome and skip a download), or decode the base64 from the transcript for a shot you already captured.
3. Compare what you see against the fundamentals and the don't-list.
4. Fix what's visibly wrong and repeat. One or two iteration loops catch what reading the code never will.

Do not declare a visual task done without having looked at the rendered result. If no browser is available in this session, say so explicitly rather than implying you eyeballed it.

Pre-ship checklist

A short gate — not an exhaustive audit. Run it before calling any screen done:

1. Squint test — the most important element is obvious within ~2 seconds.
2. Contrast — body ≥4.5:1, large/UI ≥3:1; placeholders counted.
3. Type — one scale, body ≥16px, measure ≤~75ch, ≤3 families.
4. Color — one accent; OKLCH; tinted neutrals; no AI-purple, no reflexive beige, no gradient text.
5. Spacing — consistent 4pt rhythm; grouping vs separation is varied, not uniform.
6. States — focus ring visible; hover/active/disabled/loading/empty/error all real.
7. Motion — motivated, <300ms, custom easing, prefers-reduced-motion honored.
8. Copy — verb+object buttons, useful errors, no em-dashes, no buzzwords.
9. Layout tells — no nested cards, no eyebrow chips, no numbered section markers, no three-equal-cards.
10. Responsive — single-column mobile works; nothing overflows or touches the viewport edge.
11. A11y — headings not skipped, images have alt text, touch targets ≥44px.
12. Image framing — for any embedded photo/portrait, the subject is framed correctly: faces/heads not cropped, not cut off by object-fit: cover or a fixed container aspect ratio. Verify against the ACTUAL rendered crop in the browser, not the source image — a center-crop that looks fine full-bleed can decapitate a portrait once it's constrained to the container. And don't reach for an optional, IRREVERSIBLE transform (a square/center crop, a destructive resize) unless the layout truly needs it: correctness of the shown subject comes before optimization, and a crop you can't undo isn't worth a marginal layout gain.
13. Eyeballed — you looked at the rendered result, not just the code.
14. E2E — if the project has an end-to-end test suite, a new or changed screen ships with (or updates) a test in it, run green — see the project's own docs for the harness.

Provenance

This skill distills three public design skills into Loom's own guidance. See the NOTICE file in this directory for full attribution: Impeccable (Apache-2.0, Paul Bakaus — itself building on Anthropic's frontend-design skill and ehmo's typecraft additions), taste-skill (MIT, Leonxlnx — including its v2 "design-taste-frontend" rewrite), and Emil Kowalski's design-engineering skill (MIT).