design-craft

---

name: design-craft description: Use when creating, revising, or critiquing frontend UI (web or mobile) where visual quality, product fit, design-system fidelity, or avoiding generic "AI-slop" output matters — e.g. "make this look good", "why does this look AI-generated", building a landing page / app screen / dashboard, or a design pass on existing UI. Turns vague design intent into a decision-first workflow: classify the surface, decide direction before code, escape default looks, build to measurable craft standards, and verify the rendered result. Portable across Claude Code and Codex.

design-craft

Good design is decisions, not vibes. When an LLM writes UI without deciding typography, color, layout, surface type, motion, and copy for this specific brief, it falls back to the highest-frequency patterns in its training data — the look people call "AI slop." Slop is not ugliness; it's the absence of decisions.

This skill makes those decisions explicit and enforces them. It is the synthesis of Anthropic's frontend-design methodology, a surface-type router, a measurable craft standard, a named anti-slop banlist, and a rendered-output verification loop.

Act as the design lead at a small studio whose clients have already rejected templated work and are paying for a distinct, justified point of view — not a neutral UI generator.

When to use

• Building any UI where it should look intentional: landing/marketing pages, app screens, dashboards, mobile UI.
• Someone says "make it look good / more polished / less generic / less AI-generated."
• A design pass / visual QA on existing UI.
• Reviewing whether a design reads as templated slop.

When NOT to use

• Pure logic/backend/data work with no visual surface.
• The user gave an exact, locked visual spec or design system to implement verbatim — then follow it exactly; do not fight their chosen style. (Still apply the craft floor and verification.)

Core principle: decide before you code

Do not open a .css/.tsx file until §0–§2 are done. Every color, type, spacing, motion, and copy choice in the build must trace back to a written decision. If you can't name why a choice is right for this brief, it's probably a default.

---

The workflow — six gates

Gate 0 — Classify the surface (router)

Before anything, classify the UI. Each type optimizes for different things; applying "marketing taste" to a dashboard produces slop, and vice-versa.

Surface	Optimize for	Decide first
marketing / brand / portfolio	distinct impression, one memorable moment	visual thesis, hero medium, one CTA, proof
functional app / product flow	task clarity, calm density, reuse	user task, workspace object, navigation, existing components
dashboard / analytics	the decision the user must make	the decision, KPI grouping, comparison grain, filters
mobile app	thumb reach, platform conventions	platform (iOS/Android), nav model, first-run vs returning
hybrid	which sections are marketing vs product	section-by-section split

If the surface is not marketing/brand/portfolio, reduce visual novelty and prioritize component reuse, information density, accessibility, and platform conventions over expressive direction. See references/fundamentals.md §"Surface playbook" for per-type good-defaults and anti-patterns.

Gate 1 — Design-decision pass (no code yet)

Write a short brief. Ground it in the subject's real world (its materials, vocabulary, audience) — that's where non-generic choices come from.

SUBJECT: concrete product/place/person/topic (pin it if the brief doesn't)
AUDIENCE: who, and their emotional register
JOB: the single job of this screen
VISUAL THESIS: what the first viewport proves in one sentence
COMPONENT SOURCE: existing design system / library, or "new"
TOKENS:
  color:   4–6 named roles in OKLCH (perceptually uniform, smoother gradients), not raw hex
  type:    display / body / utility roles (2+ real typefaces, chosen not defaulted — never Inter/Geist by reflex)
  layout:  symmetric vs asymmetric (by product character) + one sentence / ASCII wireframe
  signature: the ONE element this screen is remembered by (if GPU/3D-driven → references/gpu-visuals.md)
ANTI-DEFAULTS: 3 specific clichés this brief must avoid

If a design system / component library already exists, use its real components and tokens before inventing primitives. Real system context is the strongest anti-slop lever.

Gate 2 — Diverge → converge

For any surface where the visual direction is open (and always for a hero / ad / signature screen): propose 3–5 genuinely different directions, each with a one-line "default-escape point." Then pick one, state why it fits the audience and task, and reject the others with concrete reasons. Single-shot generation converges to the default; divergence-first escapes it.

If the brief already pins the direction, skip divergence and implement it faithfully.

Gate 3 — Anti-default critique (before building)

Run the design plan against the named-cliché banlist. Open references/banlist.md and check every category (layout, cards/components, color/effects, type/copy, icons/media, motion). Revise anything that:

• could appear on any SaaS landing page,
• uses cards / gradients / badges / numbers / icons as decoration only,
• has a signature not drawn from the subject,
• has copy generic enough to belong to any product.

The three AI default looks — treat as blacklisted unless the brief explicitly asks or the subject truly grounds them:

1. warm cream (~#F4F1EA) + high-contrast serif + terracotta accent
2. near-black + one acid-green / vermilion accent
3. broadsheet: hairline rules, zero border-radius, dense newspaper columns

These are defaults, not choices — they show up regardless of subject. Don't spend a free axis on one. (Note: they now recur so often they are themselves slop.)

Reframe technique when stuck: turn a component into a medium and remove the affordance — e.g. "the ad is not a card; it's an editorial insert / a sponsor line," "onboarding is not a modal; it's a journey scene." Removing the default affordance forces a real solution.

Gate 4 — Build

Implement only the chosen direction, deriving every value from the tokens.

• Spend boldness in one place. The signature is the one memorable thing; keep everything around it quiet. Then remove one accessory (Chanel's rule) — cut decoration that doesn't clarify hierarchy, brand, or task.
• Copy is design material, not decoration. Name things by what the user controls, not how the system is built. Active voice; a control says what happens ("Save changes", not "Submit"); an action keeps its name through the whole flow (button "Publish" → toast "Published"). Treat empty/error states as direction, not mood.
• Meet the measurable craft floor — body ≥16px, text contrast ≥4.5:1, line length 45–75ch, 4/8 spacing scale, motion via transform/opacity only. Full numbers in references/fundamentals.md; keep it open while building.
• Watch CSS specificity: type-selectors (.section) vs element/utility selectors can cancel each other's padding/margin — a common generated-CSS bug.

Gate 5 — Verify the rendered result

Rendered output is the truth, not the source. Before reporting done, open references/verification.md and run the browser gates: screenshot at desktop/tablet/mobile, then a screenshot-based self-critique (5-second test, hierarchy, type, color, spacing, motion, and a slop-tell scan).

Blocking rule: if the critic finds ≥2 slop tells, any accessibility hard-fail (contrast, focus, target size), mobile horizontal scroll, or missing empty/loading/error states → do a targeted rework pass (max 2) before finishing. If a fix needs a concept change, generate 3 revised concepts first.

Guard against over-correction: weird/maximalist/glassy/luxury-editorial escapes are also slop. Keep the brief's tone; take exactly one real risk.

---

Measurable fundamentals (summary)

Keep references/fundamentals.md open during Gates 1 and 4. The non-negotiable numbers:

• Type: body ≥16px (dense UI ≥14px); line length 45–75ch (never >90ch, CJK ≤40em); body line-height 1.45–1.65; display vs body are different, justified faces; no global negative letter-spacing.
• Color: define in OKLCH (perceptually uniform); text contrast ≥4.5:1 (large ≥3:1; AAA target 7:1); UI borders/focus/icons ≥3:1; 60-30-10, one saturated accent reserved for action/status; dark mode is redesigned, not inverted; color is never the only carrier of state.
• Spacing/layout: 4/8 scale (0,2,4,8,12,16,24,32,48,64,96); intra-group spacing ≤ ½ inter-group; cards only for repeated entities / choices / modals / tool frames — never wrap a hero or a page section in a card; no horizontal scroll at 360px.
• Hierarchy: in 5 seconds a screenshot answers "what is it, for whom, what next?"; emphasis order = position/size > spacing > weight > color > motion; one job + one headline per section; controlled CTA count.
• Motion: transform/opacity only; durations micro 50–100 / short 150–250 / medium 250–400ms; different easing for enter/exit/move; prefers-reduced-motion removes non-essential motion; every motion explains hierarchy, state, causality, or story.

---

Scaling up: one page → system → test

A single screen is Gates 1–5. Beyond that, three levels separate custom from templated:

• Page — the six gates above, for one surface.
• System — for multi-page work, keep a persistent design.md (the visual system: OKLCH palette, type roles, spacing/radius/motion tokens, anti-patterns) separate from project context (CLAUDE.md / AGENTS.md). Any agent should grasp the visual system from design.md alone. Put a line at its top telling agents to append newly-decided values, so it compounds each session. This is what stops page 2 (dashboard/auth) from drifting off the landing page's style.
• Test — turn design.md's anti-patterns into programmatic checks and run them before building (see references/verification.md → Design TDD). Structured, machine-readable feedback beats "looks good."

---

Using this skill in Claude Code vs Codex

The core workflow above is host-neutral markdown — it runs identically in both. Host specifics:

Claude Code

• Invoke with /design-craft or let it auto-trigger from the description.
• For visual verification (Gate 5), use claude-in-chrome / a dev server / Playwright screenshots when available.
• Large tables live in references/*; they load only when you open them (progressive disclosure).

Codex

• Codex may not auto-select this skill. When doing UI work, read this SKILL.md first, then inspect the project, then implement — or invoke via /goal pointing here. For durable per-repo defaults, mirror the craft floor into AGENTS.md.
• To make it discoverable, place a copy at ~/.agents/skills/design-craft/SKILL.md or .agents/skills/design-craft/SKILL.md.
• Gate 5 verification: use whatever screenshot tool the session has (Playwright, a headless Chrome --screenshot, etc.).

Guarded dependencies — name and check before relying on any: browser available? dev server? Figma MCP? image generation? If a verification tool is unavailable, say "visual verification not performed" rather than assuming it passed.

Reference files (open when the gate says so)

• references/fundamentals.md — typography/color/spacing/hierarchy/layout/motion as checkable rules + numbers, and the per-surface playbook. Open at Gate 1 and Gate 4.
• references/banlist.md — the extended AI-slop catalog (7 categories) + escape techniques + over-correction guard. Open at Gate 3.
• references/verification.md — browser render gates, the screenshot self-critique prompt, the quality-gate table, the rework loop, and Design TDD (anti-patterns→tests, Playwright/Vizzly, living external audit). Open at Gate 5.
• references/gpu-visuals.md — WebGL / shader / 3D (R3F + Next.js) as a deliberate signature: when to use, the 2–3 color restraint that makes 3D carry emotion, the Lenis+GSAP scroll-camera pattern, and non-negotiable perf / reduced-motion / SEO guardrails (+ "WebGL slop" guard). Open at Gate 1/4 when the signature is GPU-driven.