Jaw Dev Uiux Design

UI/UX Design: Intent Discovery, Patterns & Product Vocabulary

Activates by change surface when:

• User's design direction is vague ("깔끔하게", "모던하게", "just make it look good")
• Building onboarding, empty state, error state, or loading state UI
• User references a product aesthetic ("Notion 느낌", "Linear처럼")
• Starting a new design system or generating a color palette
• Choosing layout patterns or navigation architecture
• Setting up favicons, product logos, or brand identity elements
• Handling logo dark mode variants, OG images, or social sharing meta

Read this before style-specific references when the user cannot articulate a clear design direction. For rendered anti-slop tell detection and implementation-level banned patterns, defer to dev-frontend/references/core/anti-slop.md. This skill owns concept/taste-level anti-slop judgment (is this direction generic or domain-wrong?).

Emoji ban (stub): no emoji as UI visual elements (STRICT). Canonical rule, scope, and exemptions: dev-frontend §5 / dev-frontend/references/core/anti-slop.md § Emoji Slop.

Role separation: This skill owns design judgment: intent discovery, information architecture, UX state meaning, typography/color/layout direction, product personality, brand vocabulary, anti-slop concept/taste judgment, and design-system decisions. dev-frontend owns implementation: HTML/CSS/components, responsive mechanics, accessibility wiring, runtime behavior, rendered tell detection, and rendered verification. After choosing the design direction here, load dev-frontend for concrete implementation.

Role boundary (canonical — identical in dev and dev-frontend): dev owns universal process, evidence, and safety rules. dev-uiux-design owns design intent, direction, and concept judgment. dev-frontend owns concrete frontend implementation and rendered tell enforcement. Anti-slop has three layers: dev = output/process hygiene (FAMILY-SLOP-01), dev-uiux-design = concept/taste judgment (is this direction generic or domain-wrong?), dev-frontend = rendered implementation tell detection and removal (FE-AI-TELL-01).

External/current design evidence: For live product-reference claims, current design-system docs, browser API behavior, accessibility guidance that may have changed, or browser-rendered source evidence, read the active search skill and follow its query-rewrite, source-fetch, and evidence-status rules. Use browser fetch/open/text/get-dom/snapshot only after candidate URLs exist.

C0/C1 work (small local patches): See dev §0.0 Work Classifier + §0.1 Patch Fast-Path before reading references.

dev is canonical: dev §0.2 Rule Classes, §3 Verification Gate, and §5 Safety Rules apply to all work governed by this skill.

Rule class note (UX-STYLE-01): Everything in this skill that expresses taste — product personalities, design-isms, preset tokens, aesthetic vocabulary — is STYLE_SAMPLE: examples to draw from, never universal requirements. Objective UX correctness (state coverage, accessibility, readability) is owned by dev-frontend §1.5 and stays STRICT/DEFAULT.

Modular References

Lazy-User Gate (UX-LAZY-01, DEFAULT — ponytail discipline applied to UX)

Design for the cognitively frugal user: users don't read, they scan; they satisfice; they will trade choice for one obvious next action. Before shipping any user-facing decision point — option, setting, step, confirmation, input field, mode — justify its existence the ponytail way, in order:

1. Do nothing: can a correct default remove this decision entirely? (A settings page is a collection of defaults you failed to choose.)
2. Delete: does the step/field earn its completion-rate cost? Every added decision point taxes conversion and comprehension (Hick's law).
3. Absorb: can the system take the complexity instead of the user (Tesler's law) — auto-detect, infer, remember last choice?
4. Demote: still needed for some users → progressive disclosure (advanced section, "more options"), never a top-level fork.

Every screen has ONE primary action. If two actions compete visually, the design has not decided what the screen is for.

Surface-conditional (do not over-apply): laziness means fewer decisions on consumer/one-shot flows (D1-D3), but fewer repeated motions on repeated-work tools (D4-D8: density, keyboard paths, batch actions — collapsing an expert's controls into wizards is the inverse failure). Route by the product-density profile first.

STRICT exemptions (never one-click away): destructive/irreversible actions, consent/privacy/legal choices, payments confirmation, and accessibility affordances are never collapsed into magic defaults.

UX State Contract (UX-STATE-01)

For onboarding, empty, loading, error, or progressive-disclosure work, answer the state meaning before styling. Deep patterns live in references/ux-states.md.

• Onboarding teaches the first meaningful action, not the whole product.
• Empty explains why the state exists and names the next action.
• Loading chooses skeleton for known structure, spinner/progress for short unknown waits, and avoids fake completion.
• Error exposes retry, recovery, or escalation; never dead-end the user.
• Progressive disclosure names what stays hidden, why, and where it becomes available.

IA Chooser (UX-IA-01)

Default navigation architecture by work shape; read references/responsive-nav.md for responsive details.

1. User Intent Discovery Protocol

When the user's design request is vague ("깔끔하게 해줘", "모던하게", "just make it look good"), do not produce generic output. Run the compact ambiguity flow (UX-INTENT-01):

1. Produce the Design Read from §2 using available signals.
2. If one decision still blocks the direction, ask ONE best clarifying fork with binary/ternary choices.
3. Proceed from the answer; if the user does not answer and the task can continue, choose the most domain-correct default and state the assumption.

Skip this section if the user provided explicit design specs or this is a ≤5-line patch.

Optional deepening: use the ladder below only when the first fork fails or the user explicitly wants guided exploration.

• Use binary/ternary choices, not open-ended questions.
• Reference known products — users recognize what they want faster than they articulate it.
• If the diagram skill is available, offer: "참고로 스타일 비교를 다이어그램으로 보여드릴 수도 있어요."
• If the user names a specific product reference, skip remaining steps and map directly via references/product-personalities.md.

Step 1 — Mood

Ask: "전체적인 분위기가 어떤 느낌이면 좋을까요?" / "What overall feeling should the product have?"

Step 2 — Lightness

Ask: "밝은 화면이 좋으신가요, 어두운 화면이 좋으신가요?" / "Light or dark background?"

Step 3 — Density

Ask: "화면에 정보가 많이 보이는 게 좋으신가요, 여유롭게 보이는 게 좋으신가요?" / "Dense or spacious?"

Step 4 — Shape

Ask: "모서리가 각진 느낌이 좋으신가요, 둥근 느낌이 좋으신가요?" / "Sharp or rounded?"

Step 5 — Viewport Priority

Ask: "주로 어떤 화면에서 볼 건가요?" / "What's the primary viewing device?"

Cross-ref: references/responsive-nav.md for canonical breakpoints and container query patterns, dev-frontend/references/core/mobile-ux.md for mobile-specific composition rules.

Step 6 — Reference

Ask: "혹시 '이런 느낌이면 좋겠다' 하는 사이트나 앱이 있으신가요?" / "Any website or app that feels like what you want?"

This single question often resolves all ambiguity. If the user names a product, map it via references/product-personalities.md.

Vague Request Disambiguation

When the user gives feedback without specifics, translate:

2. Design Read (MANDATORY for new pages, components, or layouts. Optional for ≤5-line patches — see dev §0.1 Patch Fast-Path.)

Before generating ANY frontend code, produce a Design Read. If the project has a DESIGN.md file, read it first — its tokens and prose override everything below.

Output format (mini DESIGN.md)

Filled-in example: references/design-read-example.md.

---
name: <project-name>
colors:
  primary: "<hex>"
  accent: "<hex>"
  background: "<hex>"
typography:
  heading: { fontFamily: <font>, fontSize: <size> }
  body: { fontFamily: <font>, fontSize: <size> }
---

Reading this as: for , with a language. <1-2 sentences: specific reference, not adjectives. "1970s lecture handout" > "modern and clean">

Do's: Don'ts:

Signals to read

1. Page kind — landing (SaaS/consumer/agency/event), portfolio, redesign, editorial, app UI, tool UI
2. Vibe words — what the user said or implied
3. Reference signals — URLs, screenshots, brands named
4. Audience — B2B procurement vs design-conscious consumer vs recruiter
5. Existing brand assets — logo, color, type, photography
6. Quiet constraints — accessibility-first, public-sector, regulated, kids

Dial Setting (MANDATORY — immediately after Design Read)

From the Design Read, derive and declare three dials before any code:

DESIGN_VARIANCE: <1-10>
MOTION_INTENSITY: <1-10>
Product density profile: <D1-D8> (see dev-frontend/references/core/product-density.md)
Reasoning: <one sentence explaining why these values match the brief>

Inference rules:

• Corporate/gov/utility → VARIANCE 2-4, MOTION 1-3, density D2-D3
• Marketing/landing → VARIANCE 4-7, MOTION 3-5, density D2-D3
• Creative/portfolio/editorial → VARIANCE 6-9, MOTION 4-7, density D1-D3
• Dashboard/SaaS/admin → VARIANCE 2-4, MOTION 1-2, density D4-D5
• "Complex" in brief → increase density profile (functional depth), NOT VARIANCE or MOTION
• "Simple" in brief → decrease all three proportionally

"복잡하다" = high DESIGN_VARIANCE is WRONG. Complexity means more features/data/flows, not more visual tricks (carousels, parallax, animations).

Dial Presets (UX-DIAL-PRESET-01, STYLE_SAMPLE)

Source: taste-skill v2, via codexclaw. Exact tuples for common use cases. Presets are authoritative specializations that may exceed the inference ranges above — Agency motion 8 exceeds the general landing 5-7 range on purpose. When a preset exists for the exact use case, use it directly and adjust from the Design Read.

Redesign arithmetic (DEFAULT). Score the existing surface with the preset rules first, then move from it:

• Preserve: variance = match existing, motion = match + 1, density = match existing.
• Overhaul: variance = existing + 2, motion = existing + 2, density = match existing.
• "Complex" in the brief increases density, never variance or motion.
• "Simple" decreases variance and motion; density stays or increases.

Clamp all arithmetic to 1-10 and D1-D8. An existing SaaS surface at 7/6/4 preserves to 7/7/4 and overhauls to 9/8/4.

Audience-first ownership (UX-AUDIENCE-01, DEFAULT)

The audience picks the aesthetic, not the model's taste. When audience signal and model preference conflict, audience wins. The presets above encode audience expectations — a public-sector audience expects trust-first restraint, an agency audience expects high variance — so reaching past a preset because the result feels plain is the failure this rule names. Override with stated rationale only.

Anti-Default Discipline

Do not default to: warm beige backgrounds, centered hero, three equal feature cards, generic glassmorphism, Inter + slate-900, card-based everything. These are LLM defaults. Reach past them BASED ON the design read.

If the brief is ambiguous, ask ONE clarifying question. Not a multi-question dump.

DESIGN.md persistence

Image-First Direction Discovery (UX-IMAGE-FIRST-01, DEFAULT)

Fires when the brief has no named ism, product reference, or concrete design direction. Generate 5 maximally different ism directions, varying layout family, palette, type stance, material, and hero grammar. Every prompt must be detailed enough to reconstruct the layout.

Compare candidates on hero composition, palette coherence, typographic voice, and density fit. Pick the winning ism, not the winning image — the render is evidence for a direction, and treating it as the deliverable is how a one-off composition becomes an accidental design system.

Then synthesize rather than crowning one candidate. Build an element ledger recording, per token, which variant did it best and why:

The synthesis IS the direction lock: it assembles the best tokens from several candidates into one coherent set, and the lock is that assembled set rather than any single render. Refine the locked direction with 2-4 variants afterwards.

When a human is in the loop they pick the ism; running unattended, state the selection reasoning in the record.

2.6 Asset Production Handoff (UX-ASSET-GEN-01, DEFAULT)

Once the concept direction is locked, asset work belongs to jaw-dev-frontend/references/core/asset-requirements.md — generation, background removal (FE-ASSET-BG-01), batching, selection, and integration. This section exists to mark the handoff rather than to duplicate it: the direction is decided here, the assets are produced there, and blurring the two is how an unlocked direction gets 40 generated images spent on it.

2.7 Icon Strategy (UX-ICON-01, DEFAULT)

Choose iconography during the Design Read, before frontend implementation. Select the domain-correct default from the density matrix; an explicit DESIGN.md iconography block overrides it. Do not ask the user unless icon direction is a material brand decision that cannot be inferred from the brief.

Routing:

1. Set the system library from the matrix and the product's visual language. Phosphor is the general default, not a universal mandate.
2. Respect an explicit user choice or an existing DESIGN.md value over the inferred default.
3. Use the library for routine system semantics — navigation, actions, status, search, disclosure, utility controls.
4. Generate a custom icon when a domain concept has no clear library glyph, when forcing an approximate glyph would reduce comprehension, or when the brief requires a brand-specific visual style. A licensed premium set is the alternative when it already supplies a coherent domain vocabulary.

Three visual layers, kept distinct:

• System icons (library) — routine interface semantics, one coherent library and weight language.
• Domain icons (custom or premium) — product concepts, categories, KPIs, services, objects that generic libraries cannot represent precisely.
• Brand icons (custom) — identity-bearing marks, characters, mascots, signature illustrations. Never substitute a generic stroke icon for a logo.

Korean consumer apps often carry more information and rely on colorful domain category and KPI icons for fast scanning. For D4-D6 Korean-first apps, prefer a restrained system library plus a coherent colored domain layer over forcing every concept into monochrome outline icons.

If the project needs persistent design tokens across sessions, save the Design Read as a full DESIGN.md in the project root. Format spec: references/design-system-bootstrap.md § DESIGN.md Format.

3. Korean Request Translation

Map common Korean design descriptors to concrete tokens. When the user uses these words, translate before implementing.

Clarifying questions per term:

• 깔끔: "Notion처럼 따뜻한 깔끔함인지, Vercel처럼 차가운 깔끔함인지요?"
• 모던: "다크 모드 + 날카로움(Linear)인지, 화이트 + 미니멀(Vercel)인지요?"
• 고급: "브랜드 고급감(Apple/Stripe)인지, 패션 럭셔리(Art Deco)인지요?"
• 심플: "기능이 적은 건지, 기능은 많지만 화면이 심플해 보이길 원하는 건지요?"

4. Quick-Match Table

Rapid lookup: user word → concrete starting point.

Font Selection Guidelines (STYLE_SAMPLE)

• Primary default: Geist (modern SaaS, Vercel ecosystem)
• Korean-first: Pretendard — strong Korean-first default (Pretendard Variable available); verify brand/product font rules before claiming any specific Korean company standard
• Warm/editorial: Outfit or Cabinet Grotesk
• Premium/luxury: Satoshi or system thin weights
• Avoid defaulting to Inter (DEFAULT) — a widely recognized AI-generated-UI tell (judgment rule, not a measured fact). Use it when the user requests it or the project already uses it.