Ima2 Front

ima2 Frontend — Domain-Correct Frontend Engineering

Setup

npm install -g ima2-gen    # install globally (Node.js >= 20)
ima2 setup                 # first-time auth (GPT OAuth recommended)
ima2 serve                 # start local server
ima2 ping                  # verify
ima2 capabilities --json   # check models, limits, providers
ima2 defaults --json       # inspect default model/reasoning

Agent bootstrap: ima2 ping first. If unreachable: ima2 serve &. If not installed: npm install -g ima2-gen && ima2 setup. Use ima2 skill path to locate the skills directory; read ../ima2-uiux/SKILL.md for design direction and ../ima2-front/SKILL.md (this file) for implementation.

Build production-grade frontend implementations from an established product/design direction. This skill owns HTML/CSS/component/runtime implementation, responsive behavior, accessibility wiring, visual verification, and frontend platform rules.

Role separation: For design judgment — typography/color/layout direction, UX decision gates, product personalities, or vague visual briefs — load ima2-uiux first. This skill implements the chosen direction; ima2-uiux makes the design decisions. Implementation anti-slop enforcement stays here; design taste/pattern judgment lives there.

C0/C1 work (small local patches): For small patches, skip the full reference chain.

Modular References

Loading references via CLI: Recommended: install skills to your agent's skill directory.

ima2 skill install --dir <agent-skill-path>   # agent provides its own path
ima2 skill install --tmp                       # ephemeral fallback

The agent determines its own skill directory (e.g. ~/.codex/skills/, ./skills/, etc.) and passes it via --dir. After install, SKILL.md and references/ are on disk — the agent reads them natively via relative paths.

Ad-hoc reading (without install):

ima2 skill front refs              # list all reference modules with line counts
ima2 skill front ref motion        # print one module (basename match)
ima2 skill front ref stacks/react  # print a nested module

Start with anti-slop.md, aesthetics.md, responsive-viewport.md, and visual-verification.md. Add domain/locale/stack references only when relevant. For C2 ordinary app screens (form/table/list/detail), crud-ui.md alone suffices; add the style references above for marketing/visual surfaces or C3+ work.

When frontend choices depend on current framework, design-system, browser API, library behavior, browser-rendered source evidence, or package/source freshness, read the active search skill and follow its source-fetch and evidence-status rules before treating external material as proof.

Verification grounding

STRICT: For render/executable artifacts (HTML, SVG, games, UI, charts), run the real renderer: headless browser, screenshot, canvas check, or equivalent. Observe the actual output yourself, fix what observation reveals, then re-run. Static parsing confirms well-formed files; it does not prove the artifact is visually or interactively correct. One clean observation is enough for unchanged state; do not re-render unchanged output just to repeat evidence.

0. Frontend Routing

Before designing or coding, classify the work:

Default rules:

• For apps/tools/dashboards, build the actual working surface first, not a marketing hero.
• For Korean-first work, read korea-2026.md and ux-writing-ko.md.
• For any soft 3D miniature, mascot, chibi, toy-like object, or character-like asset, read soft-3d-asset-gates.md.
• For product/brand/object/place/person pages, use concrete visual assets in the first viewport.
• For finance, government, B2B, admin, auth, security, and developer tools, keep visual warmth restrained and subordinate to clarity.
• Every user-facing decision point must justify its existence — defaults first, one primary action per screen, choices demoted to progressive disclosure (ima2-uiux UX-LAZY-01 owns the gate).
• For text-heavy surfaces (landing, marketing, editorial, public service), apply typography wrapping defaults — see typography-wrapping.md. Dashboard table cells are excluded.

1. Component Identification

When the user describes UI in vague terms (e.g. "접히는 거", "팝업 같은 거"):

1. Recommend the best-fit component with reasoning: <Name> — <what it does, why it fits>
2. Confirm, then proceed

If the user already names a specific component, skip this step. Reference: component.gallery/components

For new React/Vue/Svelte/Next UI source files, prefer .tsx or typed component files when the repo supports TypeScript. Inherit dev TypeScript strict-compatibility rules. If frontend structure is unclear, read existing source-of-truth docs first, then document pages, components, routes, state stores, and build commands in the repo's existing docs before broad implementation.

1.5 Objective Gates vs Style Samples

Two different kinds of rules live in this skill (see the work classifier):

• Objective UX gates (STRICT/DEFAULT) — accessibility baseline (§7, §11), state coverage (loading/empty/error/permission), keyboard operability, visible focus, contrast. Missing these are review findings.
• Style direction (STYLE_SAMPLE) — design direction intake (§2), aesthetics, density profiles, product personalities, preset tokens, and the concrete values in §4-§5 (palettes, font choices, pixel max-widths). These illustrate acceptable choices; they are NOT requirements, must not override an existing design system (Design System Detection stays MANDATORY), and must never be enforced as universal taste (UX-STYLE-01).

2. Design Direction Intake

When the user cannot articulate a clear design direction, load ima2-uiux to discover intent and choose a direction before implementing here.

Before coding, commit to a domain-correct direction:

• Purpose: What problem does this interface solve? Who uses it?
• Surface: Is this a working tool, dashboard, public service, AI workflow, game, landing page, or editorial surface?
• Tone: Pick a specific direction. For product tools this often means quiet, dense, trustworthy, and fast rather than loud.
• Constraints: Framework, performance budget, accessibility requirements.
• Signature: What ONE thing will make this unforgettable? (the signature moment; supporting scroll reveals may exist alongside it per motion.md FE-MOTION-BUCKET-01)

When user intent is vague ("깔끔하게", "모던하게", "just make it look good"), read the ima2-uiux skill and run the User Intent Discovery Protocol before making routing decisions. If the user cannot answer these questions, use the ima2-uiux skill's structured preference elicitation flow. Offer product references ("Notion 느낌? Linear 느낌?") and visual comparisons.

Concept pass before code (stub — canonical: ima2-uiux §2.5 UX-CONCEPT-GEN-01): for a C2+ new/redesigned expressive or brand-visible surface (page, hero, key chrome like a top bar) with open design direction — probe ima2 status, attempt ima2 serve if down, then run ima2 models --kind image and configure ima2 defaults set image <lane>/<model> before any bare ima2 gen; use the native image generator only as a true fallback — then generate 5 highly specific candidate mockups of ONE locked concept, then SYNTHESIZE the best elements across all 5 (NOT pick a single winner) into DESIGN.md and implement from that synthesis — do not start coding the layout blind.

Intentionality over intensity. Bold maximalism, refined minimalism, dense utility, and friendly consumer UI can all work when they match the domain.

3. Baseline Configuration

Adjust these dials based on what's being built. Present to user if unclear.

After Design Read, set dials per ima2-uiux §2 Dial Setting.

Product density profile (D1-D8 in references/product-density.md) sets component class; VISUAL_DENSITY (1-10) sets spacing within that class. These are orthogonal axes.

Adapt dynamically based on user requests. Dashboard → density up. Portfolio → variance up. Data tool → motion down. Korean app/tool surfaces usually need higher density and clearer hierarchy, not oversized hero text.

4. Implementation

Read references/aesthetics.md for full guidelines. Summary:

• Typography: Use domain-appropriate typography. For Korean-first UIs, prioritize CJK-safe stacks before Latin display fonts. Apply text-wrap: balance on all headings AND short descriptors (hero subtitle, card description, caption — anything 1-3 lines). Use text-wrap: pretty only on body paragraphs (4+ lines). pretty has no effect on short text and will leave Korean orphans like "합니다." or "화." on a line alone. See typography-wrapping.md for full rules.
• Color: Max 1 accent. Use neutral bases (Zinc/Slate) with singular high-contrast accent — avoid purple-on-white.
• Layout: Match the product surface. Avoid centered-card/hero patterns in repeated-use tools.
• Motion: See references/motion.md. One signature moment + a few supporting reveals > 10 scattered effects; landing-bucket floor/ceiling per FE-MOTION-BUCKET-01.
• Assets: Use screenshots, product images, diagrams, charts, illustrations, generated bitmaps, or soft 3D only when they add product meaning. When a real bitmap is needed (icon, hero, illustration), generate it with ima2 — probe ima2 status, attempt ima2 serve if down, inspect ima2 models --kind image, and set ima2 defaults set image <lane>/<model> before using bare generation — falling back to the native imagegen tool only when ima2 is truly unavailable; never ship a placeholder. ima2 is preferred because it supports reference images, multi-candidate generation (-n N, multimode, independent CLI parallel — see asset-requirements.md FE-ASSET-PARALLEL-01), a backend-selectable prompt builder, raster-to-vector tracing (ima2 vectorize and Canvas trace — see asset-requirements.md Raster vs real vector), session style sheets, provider routing (GPT/Grok/Gemini — see asset-requirements.md FE-ASSET-PROVIDER-01), variant selection with element-ledger synthesis (asset-requirements.md FE-ASSET-SELECT-01), cutout asset background strategy (asset-requirements.md FE-ASSET-BG-01), and video (ima2 video — see motion.md FE-MOTION-VIDEO-01) for motion assets. For parallel generation, monitor with ima2 ps --json and cancel unwanted jobs with ima2 cancel <id>. Write very explicit long prompts (subject, composition, palette, lighting, style, aspect) per asset-requirements.md; prefer real/generated image or video assets over CSS gradient washes. Read any design reference or captured screenshot back into context with view_image before matching it. Third-party captures follow reference-capture.md (analysis-only, provenance manifest).
• Visual verification: after UI changes, exercise the flow per visual verification (screenshot -> view_image) — browser:control-in-app-browser on the dev server, screenshot, view_image — instead of claiming visual correctness from code alone.

Cutout Asset Generation (FE-ASSET-BG-01 surface — STRICT)

GPT Image 2 cannot produce transparent backgrounds. Requesting "transparent background" or "PNG with alpha" yields checkerboard artifacts or solid fills. Every cutout asset (icons, product shots, 3D objects, illustrations, stickers, UI elements that float over arbitrary backgrounds) MUST use the solid-background- then-remove pipeline. Full rules and recipes: references/asset-requirements.md § Asset Background Strategy.

Quick reference — generation template:

# Reflective/metallic/glass subjects → PURE BLACK bg
ima2 gen "3D render of [subject], [material/style details], [composition]. \
  Floating on a PURE SOLID BLACK background. The background must be 100% flat \
  pure black hex #000000. No checkerboard, no transparency pattern, no gradient, \
  no floor plane, no shadow, no vignette, no ambient glow on the background." \
  --model oauth/luna --quality high --size 1024x1024 --mode direct -o asset.png

# Dark/opaque subjects → PURE WHITE bg
ima2 gen "[subject description], centered, floating. PURE SOLID WHITE background \
  hex #ffffff. No shadow, no gradient, no surface, no reflection plane." \
  --model oauth/luna --quality high --size 1024x1024 --mode direct -o asset.png

# Known destination color → match it exactly
ima2 gen "[subject description], centered. PURE SOLID background hex #[target]. \
  No gradient, no texture, no shadow." \
  --model oauth/luna --quality medium --size 512x512 --mode direct -o asset.png

Quick reference — CSS removal (zero post-processing):

For programmatic removal (build pipelines): sharp, ImageMagick, or rembg. For interactive cleanup: ima2 Canvas Mode. For targeted fix: ima2 edit.

5. Anti-Slop Enforcement

Rule classes (dev §0.2): items below are DEFAULT — deviate with a stated reason; concrete values and palettes are STYLE_SAMPLE (§1.5); the emoji-as-UI-icon ban is the only STRICT item.

Read references/anti-slop.md for full rules. Key standards:

Hero discipline (FE-HERO-01)

• First viewport must fit: hero content leaves a hint of the next section on mobile and desktop.

• Keep hero copy to ~4 text elements max: headline, subhead, primary CTA, one proof/context line.

• Do not put trust strips, pricing teasers, feature bullets, or mini dashboards inside the hero.

• Logo walls belong below the hero, not as hero filler.

• Plan font scale with image/product scale so neither crushes the other.

• Treat unexamined default typography as a slop signal. Choose a domain-appropriate stack; Korean-first UI should use CJK-safe fonts and system fallbacks deliberately.

• Gradient budget (FE-GRADIENT-01): gradient soup is the 2026 #1 anti-slop signal; max 1 ambient gradient per viewport and no gradients on 3+ sibling cards — see anti-slop.md § Gradient Budget

• One-note theme ban (FE-ONENOTE-01): full-page single-hue dark washes (terminal green, cyber cyan, CRT amber) are the current dark-mode tell — see anti-slop.md § One-Note Theme Ban

• No self-describing meta copy (FE-METACOPY-01): UI text must explain the product/user job, never narrate the mockup, layout, responsive behavior, or agent process — see anti-slop.md § Self-Describing Meta Copy

• Use neutral or intentional color palettes — purple gradients on white are now the old tell; gradient overuse and one-note single-hue themes are the current tell

• Use asymmetric or purposeful layouts — centered-everything reads as template

• Vary card sizes, spans, and groupings — equal 3-card grids read as generic

• Bento composition (FE-BENTO-01): bento grids must read as one interlocking slab with aligned row edges, a dominant cell, content-weighted spans, and no orphan tail — see layout-discipline.md § Bento Composition

• Avoid oversized bold hero text inside tools, dashboards, admin, finance flows, and public services

• Hero composition (FE-HERO-SPLIT-01): never build a split hero (left bold headline + right boxed screenshot/mockup card) unless the user explicitly requests one — the product visual is the stage (full-width, background, or interactive demo), never a right-column card; paid-conversion LPs are the one context to propose it — see layout-discipline.md § Hero Composition Grammar

• Avoid asset-free UI: abstract blobs/gradients do not replace real visual evidence

• Avoid generic soft 3D icon packs; soft 3D must be semantic, brand-consistent, and restrained

• NEVER use emoji as UI visual elements (feature icons, card icons, section markers, buttons) — emoji in production UI is the #1 AI slop signal. Use SVG icons (Lucide/Phosphor/Heroicons). See anti-slop.md § Emoji Slop

• Warm beige/cream backgrounds with brass/clay accents are banned as defaults for premium-consumer briefs — see anti-slop.md § Premium-Consumer Palette Ban

• Layout monotony (same family repeated, 3+ zigzag sections, overused eyebrows) — see references/layout-discipline.md

• Color, shape, and theme must be locked per-page and audited before shipping — see references/consistency-locks.md

• Use off-black (#0a0a0a, #111) — pure #000000 lacks depth

• Responsive enforcement: every multi-column section must declare its mobile/tablet collapse behavior — "it'll work at mobile" is not a plan. See responsive-viewport.md

• Page containment required: max-w-[1400px] mx-auto or equivalent wrapper. Content stretching to viewport edges on wide monitors is a layout bug

• Mobile is a different product: section composition, CTA placement, and interaction model change on mobile — it is NOT just "desktop stacked vertically." See mobile-ux.md

• Use realistic, specific names and brands in placeholder content

• Write original copy — avoid "Elevate", "Seamless", "Next-Gen" and similar clichés

• Treat uncontrolled heading line breaks (orphaned single word, no text-wrap, no max-width in ch) as a slop signal — see typography-wrapping.md

• Treat short descriptors (hero subtitle, card description, caption) using text-wrap: pretty instead of balance as a slop signal — pretty does nothing on 1-3 line text, especially Korean

• Treat Korean orphan fragments ("합니다.", "화.", "입니다." alone on a line) as a slop signal — always verify Korean text breaks at target viewports

• Treat generic stroke icons as brand logo substitutes as a slop signal — use actual brand SVGs from Simple Icons, SVGL, or press kits. See brand-asset-sourcing.md

• When NO design brief exists, do not invent a generic default: apply the domain-gated no-brief kit owned by ima2-uiux §1 UX-DEFAULT-ISM-01 and state the assumption

Do not ship these tells (FE-AI-TELL-01)

Version labels in heroes, numbered eyebrows, middle-dot overuse, duplicate image reuse, monospace uppercase card labels, fake social-proof headers, decorative scroll cues, weather/status strips with no product purpose, photo-credit captions in UI chrome, and generic "trusted by teams worldwide" claims are AI-default tells. Full catalog: references/anti-slop.md + references/layout-discipline.md.

6. Performance Guardrails

• Animate transform and opacity only — layout properties (top, left, width, height) cause jank
• Grain/noise filters → fixed pseudo-elements only, keep off scrolling containers
• will-change sparingly — remove after animation completes
• Z-index only for systemic layers (navbar, modal, overlay)
• Memoize perpetual animations in isolated components

Browser Connection Limits

Rules:

• Never open >2 SSE/WebSocket connections to the same origin from one page
• Use connection multiplexing (single WebSocket with channel/topic routing) over multiple connections
• If >6 parallel requests needed: use HTTP/2, batch API endpoints, or domain sharding (last resort)
• Preflight OPTIONS requests count against the connection limit; consolidate CORS-heavy calls

Banned:

• Opening unbounded WebSocket connections per component instance
• Polling from multiple components independently (centralize into one subscription, fan out via state)
• Creating new SSE connections on every remount without cleanup

7. Accessibility Baseline

• Semantic HTML (<button>, <nav>, <main>)
• Keyboard navigation for all interactive elements
• WCAG AA minimum (4.5:1 normal text, 3:1 large text)
• Visible focus indicators (focus-visible:ring-2)
• prefers-reduced-motion support
• Skip link or equivalent bypass for repeated navigation
• Focus must not be hidden by sticky headers, sticky bottom bars, sheets, or overlays
• Icon-only buttons need accessible names (aria-label, visible text, or labelled-by)
• Charts, status messages, loading progress, and AI streaming states need screen-reader labels or live regions where appropriate
• Do not encode meaning by color alone
• Modals, menus, comboboxes, bottom sheets, and command palettes must have a complete keyboard path
• Stress-test Korean long labels and screen-reader names; clipped Hangul is a failure
• Pointer targets follow WCAG 2.2 AA target-size rules; 44×44px is a conservative product baseline, not the only legal minimum

A11y polish (FE-A11Y-POLISH-01)

• CTA text fits on one line at target breakpoints; if it wraps, shorten the label or change the layout.
• Inputs need visible boundaries against their background in default, focus, error, and disabled states.
• Duplicate CTA intent on the same screen should merge or clearly differ by outcome.
• Button contrast is checked during visual review, not left to palette intent.

8. Custom Hooks

Create a custom hook only when it owns reusable behavior, not just because code is a few lines long.

Good hook candidates: subscription lifecycle, reusable async state machine, form-field behavior shared across components, media/query/observer integration, keyboard/focus behavior, external store wrapper.

Avoid hooks that are merely thin aliases for useState, useToggle, useDebounce, or one-off component logic unless the repo already standardizes them.

Hook rules:

• The hook name describes behavior, not implementation
• Inputs are explicit and stable; return shape is small
• Side effects are justified by an external system; cleanup is correct
• Dependencies are honest; use useEffectEvent for non-reactive callbacks inside Effects
• Do not hide server state, router state, or form ownership inside a generic hook

9. React Performance

Default performance strategy: keep components pure, keep state local, classify state ownership correctly, use server rendering/caching boundaries, split expensive client islands, measure before memoizing.

If React Compiler is enabled, remove defensive memoization unless measurement or semantics justify it. Split at route boundaries and heavy components (charts, editors, 3D).

10. Form Handling

For simple forms, use controlled components with schema validation (Zod). For complex forms (multi-step, dynamic fields), use react-hook-form + Zod resolver. Always show field-level errors with role="alert".

11. Accessibility Quick-Wins

Beyond the baseline (§7):

• Focus management: trap focus in modals, restore on close, handle Escape
• Arrow keys navigate lists and menus; Enter/Space activate buttons and links
• Tab order follows visual flow
• aria-expanded, aria-haspopup, aria-activedescendant on composite widgets
• Test with screen reader and keyboard-only navigation

12. 2026 Frontend Platform Rules

Use this section when modernizing or creating React/Next/Vite frontends. Prefer project conventions first.

React 19.2+

• Activity: Use <Activity> for state-preserving hidden UI (tabs, drawers, route shells). Do not use for security hiding or active subscriptions.
• useEffectEvent: For non-reactive logic inside Effects that needs latest props/state without resubscribing. Never call during render or pass to children.
• Partial Pre-rendering: Design pages as static shell + explicit dynamic holes + Suspense boundaries. No Date.now(), Math.random(), or request-specific data in the pre-rendered shell.
• React Compiler: Do not cargo-cult memo/useMemo/useCallback. Measure first unless referential stability is semantically required.

Next.js 16

• Turbopack is default. Do not add custom webpack config unless proven unsupported.
• Cache Components (cacheComponents: true): dynamic rendering is default; cache only what you explicitly mark with use cache + cacheLife + cacheTag.
• Never cache user/session-specific data without explicit user-scoped cache key.
• Server Actions: validate input server-side, authorize against the resource, revalidate affected cache tags.

Modern CSS

Prefer native CSS before JS layout observers or animation libraries:

• Container queries for component-level responsive layout (not viewport)
• :has() for parent/sibling state selection — keep selectors narrow
• CSS nesting for modularity — keep shallow, avoid specificity tunnels
• Subgrid when nested content must align to outer grid
• View Transitions for meaningful state continuity — respect prefers-reduced-motion
• Modern units: dvh/svh/lvh over 100vh, logical properties over left/right
• Tailwind v4: CSS-first configuration, use theme variables over hardcoded values

Build Tools

• Vite 8 (verified 2026-07-02): Rolldown/Oxc is the integrated default bundler (rolldown-vite is only a Vite 7 migration bridge). Node 20.19+/22.12+; Baseline target Chrome/Edge 111, Firefox 114, Safari 16.4. Detect Vite 7 vs 8 before editing config.
• Agent-visible runtime diagnostics (DEFAULT): prefer dev servers that surface browser/runtime errors to the CLI/agent — Vite 8 forwards browser console to the dev server (auto-activates for coding agents); Next 16 ships DevTools MCP. Wire these before debugging rendered behavior.
• Do not introduce Webpack-era config unless the existing app is already Webpack-bound

State Classification

Before adding state, classify it:

Rules: Do not store derived state just to sync with Effect. Do not put server state in Zustand. Do not put URL-shareable state only in component state. Keep optimistic state reversible.

Design System Detection (MANDATORY — before creating tokens)

Before inventing design tokens, check:

1. Does the project have an installed design system? (grep -r "material-ui\|@mui\|carbon-components\|@carbon\|@fluentui\|govuk-frontend\|uswds" package.json)
2. Does the project have existing tokens? (find . -name "tokens.*" -o -name "theme.*" -o -name "design-system*")
3. Does the brief name a specific design system?

If YES to any: use the official package. Do not recreate CSS by hand.

If NO: proceed with ima2-uiux/references/design-system-bootstrap.md.

shadcn/ui and AI-Assisted UI

• Inspect existing installed components before adding new ones
• Use project's components.json, aliases, tokens, and registry conventions
• Do not hallucinate design-system components; verify against local source
• Remove demo-only copy and unused variants

For AI-native interfaces (chat, agent, copilot), design explicit states: empty → prompt ready → submitted → streaming → tool call → result → complete → feedback. Never fake streaming, citations, or tool calls.

13. Error Boundaries

React Error Boundary pattern:

• Wrap each major section (not the entire app) in an Error Boundary
• Error boundary renders: friendly message + retry button + report link
• Log error to monitoring service (Sentry, etc.) in componentDidCatch
• Never show stack traces to end users

Error state hierarchy:

1. Field-level: inline validation message
2. Form-level: summary at top of form
3. Section-level: Error Boundary with retry
4. Page-level: error.tsx / error page
5. App-level: root Error Boundary → offline/crash page

14. Pre-Flight Checklist

Checklist items apply to production surfaces (the work classifier shared definition); prototypes, spikes, and internal demos are exempt unless the user asks for production polish.

Before delivering:

• Domain-correct direction chosen and committed
• Product surface, locale, density, asset need, soft 3D gate, and motion intensity classified
• Anti-slop patterns enforced (§5)
• Hero discipline enforced: viewport fit, copy count, no in-hero trust/pricing/feature clutter (§5)
• Required assets are real, semantic, rendered, and not generic decoration
• Korean-first UI follows CJK typography and Korean UX writing rules
• Soft 3D/miniature/character assets pass domain and semantic gates
• Mobile layout collapse guaranteed with per-section-type rules (see layout-discipline.md § Responsive Transforms)
• Full-height sections use min-h-[100dvh] not h-screen
• Page containment: max-w-[1400px] mx-auto wrapper present (see responsive-viewport.md)
• Tested at 768px (tablet) and 1024px (split-screen) in addition to mobile/desktop
• Touch targets ≥ 44px on mobile; no hover-only interactions (see mobile-ux.md)
• Responsive images use srcset/sizes or <picture> for art direction (see responsive-viewport.md)
• Safe area padding for notched devices: env(safe-area-inset-*) on fixed elements
• Loading, empty, and error states provided
• State classified before adding store/Context/Effect/cache (§12)
• Effects sync with external systems; derived state is not Effect-synced
• Container queries considered before viewport-query or JS layout workarounds
• View transitions respect reduced motion
• shadcn components follow local registry and token conventions
• AI UI states are honest: no fake streaming, citations, or tool calls
• Forms validate with schema and show field-level errors (§10)

…(truncated)