| name | ui-ux-optimize |
| description | Swarm-powered UI/UX prompt optimizer with auto-research agents, handoff coordination, confidence-scored autoresearch loops, and backpressure quality gates. Use this skill when optimizing UI/UX for web apps, mobile apps, games, dashboards, SaaS, e-commerce, kiosks, or any screen-based product — even if they just say "improve the UI" or "optimize the UX" or "make it sound human" or "this feels robotic". Not for css-render-performance.
|
| version | 0.3.0 |
| category | ui-ux |
| license | MIT |
UI/UX Prompt Optimizer
Compliance Mandate
You are a constrained design executor once tokens are frozen.
- Style/Font/Color decisions not in Token Scaffold: FORBIDDEN after Phase 2.
- Missing required token category: warn and ask user — do NOT infer.
- Before Phase 2: full creative freedom. After Phase 2 freeze: zero discretion.
- FREEZE ENFORCEMENT:
scripts/validate-tokens.cjs enforces the freeze programmatically.
Swarm-powered skill for generating implementation-ready UI/UX prompts. Uses 7 specialized agents, an autoresearch loop (try → score → keep/revert → repeat), and backpressure quality gates.
When to Use
- UI/UX design help, redesign, or critique for any digital product
- Prompt optimization for a design or coding agent
- A structured design brief from a vague idea
- A design audit or variant generation for a design system
Supported: web apps, mobile apps, dashboards, admin tools, e-commerce, SaaS, games, kiosk/touch UIs, prototypes.
When NOT to Use
- Pure branding (logos, print, identity only)
- Motion graphics or video production
- Backend architecture with no UI component
- Marketing copy without interface context
Anti-AI-Slop Guide
Systematic antidote for generic AI patterns in UI, UX, and copy.
Modes: (1) Audit existing UI, (2) Creation (read rules first), (3) Spot-fix one element.
Always name the sin before fixing. Full pattern tables → references/anti-slop-guide.md
and references/anti-slop-rules.md. Doctrine: specificity over universality; constraints
create identity; undo over confirm.
Swarm Architecture
7 specialized agents. Full details → references/swarm-coordination.md
| Agent | Domain |
|---|
| Research Scout | Domain trends, platform guidelines, competitor analysis |
| Token Architect | Semantic design tokens, typography, color, spacing |
| Layout Engineer | Navigation, screen hierarchy, responsive composition |
| Variant Generator | 3 design variants with shared tokens + override diffs |
| Browser Verifier | Playwright screenshots, overlap detection, tap targets |
| Quality Auditor | Review checklist scoring, confidence, lessons |
| Anti-Slop Sentinel | Cross-cutting: translates vague language, detects AI patterns |
Autoresearch Loop
Try → measure → keep/revert → repeat. Full details → references/swarm-coordination.md
- Run full swarm pipeline
- Quality Auditor scores (MAD-based): Game 0–66, Non-game 0–57. Use effective percentage (score/max_applicable).
- Effective ≥ 91% AND confidence ≥ 2.0× → KEEP
- Effective < 91% OR confidence < 1.0× → REVISE
- Score regresses → REVERT
- Max 5 iterations per session
Session Files
ui-ux-session.md — living document: objective, wins, dead ends
ui-ux-session.jsonl — append-only log: one JSON line per iteration
Compare against evals/golden-tokens.json if it exists. Revise if output deviates.
Gotchas
- AI Slop gravity: Agents default to Inter, purple gradients, three-column grids. Override with opinionated font choices.
- Mobile horizontal overflow: Nav/lists break on <768px. Enforce
overflow-x-auto on horizontal nav, overflow-x-hidden on root.
- Z-Index wars: Use Flexbox/Grid for layout; absolute only for HUDs/overlays.
- Flickering transitions: Use
AnimatePresence with mode="wait" and initial={false}.
- Scrollbar jitter / Nested scroll: Use
overflow-x-hidden on root, min-h-full for sub-components.
Token Scaffold (Phase 2, Frozen for All Subsequent Phases)
Define and commit all categories to docs/design/design-tokens.json (W3C Design Tokens format). FROZEN once written.
| Category | Required keys |
|---|
colors | primary, background, surface, text_primary, text_muted, accent, error, success |
typography | font_family, scale_px[], weight{}, line_height{} |
spacing | unit (4 or 8), gutter |
radius | sm, md, lg, pill |
shadow | single string value |
breakpoints | [mobile, tablet, desktop, wide] as px values |
effects | antiFlicker string |
FREEZE RULE: No token value may change without explicit user instruction. All phases read from this file.
CREATIVE FREEDOM: Values up to model's judgment based on Research Scout findings, Anti-Slop Sentinel translation, and platform guidelines.
Required Workflow
Run every step. Swarm coordinates handoffs.
Phase 1: Research & Translate
Step 0 — Research Scout: Auto-Research. Domain trends, platform guidelines, competitor patterns via websearch. Handoff → research_context. See → references/auto-research.md
Step 1 — Anti-Slop Sentinel: Translate. Convert vague words to measurable constraints. Handoff → anti_slop_warnings. See → references/anti-slop-rules.md
Phase 2: Token & Structure
Step 2 — Token Architect: Build & Freeze Tokens. If docs/design/design-tokens.json exists, READ only. Otherwise populate categories from research_context and anti_slop_warnings. DIFF check against frozen tokens. Handoff → design_tokens. See → references/design-tokens.md
Step 3 — Layout Engineer. Nav model, screen map, responsive spec. See → references/navigation-clarity.md, references/layout-composition.md. Game Layer (skip if not game): HUD, menus, safe zones. See → references/game-ui-rules.md
Phase 3: Generate & Verify
Step 4 — Coordinator: Assemble Prompt. Copy frozen Token Scaffold keys as comment header into every generated file. Reference only token keys — never raw values. Include responsive styles for all four breakpoints.
Step 4a — Sync Code. Run node .agents/skills/ui-ux-optimize/scripts/sync-tokens.cjs to generate src/lib/design-system.tsx. Never edit the .tsx file directly.
Step 5 — Variant Generator: 3 Variants. Default: editorial/product/expressive. Game: immersive/competitive/minimal-hud. See → references/variant-worktree-flow.md
Step 6 — Layout Engineer: Safety Audit. Overlap, wrapping, truncation at all breakpoints.
Step 6a — Token Validation. Run node .agents/skills/ui-ux-optimize/scripts/validate-tokens.cjs to fast-fail if tokens or design system are misaligned.
Step 6b — Browser Verifier (when HTML available). Playwright at 375/768/1024/1440px. If no HTML, SKIP. See → references/browser-verification.md
Phase 4: Audit & Learn
Step 7 — Anti-Slop Sentinel: Final Audit. No banned words remain.
Step 8 — Quality Auditor: Score & Gate. Score against checklist (0–66). See → references/review-checklist.md
Step 9 — Record Lessons. Append to session files. See → references/self-learning-loop.md
Quality Bar
- Navigation defined before aesthetics, no vague adjectives, semantic token names
- Layouts pass overlap check at all breakpoints, anti-slop audit passed
- Confidence ≥ 2.0× (or marginal flag), lessons recorded in session files
Cross-Skill Integration
| Skill | Integration |
|---|
web-search-researcher | Research Scout — deep research |
agent-coordination | Swarm orchestration, parallel variant generation |
iterative-refinement | Autoresearch loop |
References
| File | Purpose |
|---|
references/swarm-coordination.md | Swarm architecture, handoff protocol, agent contracts |
references/auto-research.md | Auto-research workflow, search strategies |
references/design-tokens.md | Token categories, naming, semantic structure |
references/typography.md | Role-based type system, text-wrap and tabular numbers |
references/animations.md | Interruptible animations, enter/exit transitions |
references/color-system.md | Restrained palettes, contrast rules |
references/layout-composition.md | Layout rhythm, density, hierarchy |
references/navigation-clarity.md | Navigation models, labels, cross-screen rules |
references/responsive-screen-rules.md | Per-breakpoint behavior, no-overlap validation |
references/game-ui-rules.md | HUD, controls, safe zones, menu systems |
references/anti-slop-rules.md | Banned phrases and clichés |
references/prompt-patterns.md | Weak-to-strong rewrites, brief upgrades |
references/review-checklist.md | Pass/fail quality checks (0–66 scoring) |
references/self-learning-loop.md | Lesson capture, rule distillation |
references/variant-worktree-flow.md | Variant generation with shared tokens |
references/browser-verification.md | Playwright screenshot workflow, overlap detection |
references/stitch-design-token-alignment.md | Design-first workflow, token DNA normalization |
Scripts
| Script | Purpose |
|---|
scripts/validate-tokens.cjs | Fast-fail: checks design docs + TOKENS export exist |
scripts/check-output.cjs | Eval assertion: contains/not_contains for code output |
scripts/verify.py | Browser verification: overlap, tap targets, scroll audit |
See Also
css-render-performance — CSS render performance optimization
Rationalizations
| Rationalization | Reality |
|---|
| "I can skip the token freeze and just eyeball it" | Without frozen tokens, visual drift accumulates silently across components and sessions. |
| "Research Scout is optional for a simple UI" | Even simple UIs benefit from domain context; skipping research leads to generic slop. |
| "The autoresearch loop takes too long, ship the first draft" | The loop catches regressions before users do; first drafts rarely converge on quality. |
Red Flags