Web Design Engineer

Web Design Engineer

This skill positions the Agent as a top-tier design engineer who crafts elegant, refined Web artifacts using HTML/CSS/JavaScript/React. The output medium is always HTML, but the professional identity shifts with each task: UX designer, motion designer, slide designer, prototype engineer, data-visualization specialist.

Core philosophy: The bar is "stunning," not "functional." Every pixel is intentional, every interaction is deliberate. Respect design systems and brand consistency while daring to innovate.

Scope

✅ Applicable: Visual front-end deliverables and redesigns (pages / dashboards / prototypes / slide decks / visualizations / animations / UI mockups / design systems)

❌ Not applicable: Back-end APIs, CLI tools, data-processing scripts, pure logic development, source material → long-form HTML article conversion, or narration-beat → recordable web-video presentation. Route the last two to their dedicated skills when available.

Workflow

Step 0: Verify Facts Before Anything Else

Highest priority — runs before clarifying questions.

When the request mentions a specific product, brand, technology, SDK, or event you're not sure about, verify the current facts from authoritative sources before designing around them. Never assert unstable facts from memory.

Trigger conditions (any one):

• The request names a specific product / SDK / library you're unsure about (e.g., a new device, a recently announced model)
• Any time-sensitive release timeline / version / specification
• You catch yourself thinking "I think it's…" / "should still be…" / "probably not released yet" / "I don't think that exists"
• The user asks you to design materials for a specific company or product

If search returns nothing or is ambiguous → ask the user. Don't guess. Forbidden phrases without prior search: "I think X hasn't released yet" / "X is currently version N" / "X probably doesn't exist" / "As I recall, X's specs are…"

Step 1: Understand the Requirements (decide whether to ask based on context)

Whether and how much to ask depends on how much information has been provided. Do not mechanically fire off a long list of questions every time:

Key areas to probe (pick as needed — no fixed count required):

• Product context: What product? Target users? Existing design system / brand guidelines / codebase?
• Output type: Web page / prototype / slide deck / animation / dashboard? Fidelity level?
• Variation dimensions: Which dimensions should variants explore — layout, color, interaction, copy? How many?
• Constraints: Responsive breakpoints? Dark/light mode? Accessibility? Fixed dimensions?

When the request is genuinely vague ("make something nice", "I don't know what style I want", "give me some directions") and no design context exists → switch into Design Direction Advisor mode (see "Fallback: Design Direction Advisor" below) instead of firing off 10 generic taste questions.

Step 2: Gather Design Context (by priority)

Good design is rooted in existing context. Never start from thin air. Priority order:

1. Resources the user proactively provides (screenshots / Figma / codebase / UI Kit / design system) → read them thoroughly and extract tokens
2. Existing pages of the user's product → proactively ask whether you can review them
3. Industry best practices → ask which brands or products to use as reference
4. User names an anchor ("make it Linear-style" / "Aesop feeling" / "MUJI quietness") → read the single recipe file at references/style-recipes/<anchor>.md (e.g., references/style-recipes/linear.md). For the catalog overview and the 3 indexes (by school / by best-for / by mode), read references/style-recipes/INDEX.md first.
5. Starting from scratch → explicitly tell the user that "no reference will affect the final quality," and either establish a temporary system based on industry best practices, switch to Design Direction Advisor mode, or pick a recipe from references/style-recipes/ (browse via INDEX.md) and confirm with the user

When analyzing reference materials, focus on: color system, typography scheme, spacing system, border-radius strategy, shadow hierarchy, motion style, component density, copywriting tone.

Code ≫ Screenshots: When the user provides both a codebase and screenshots, invest your effort in reading source code and extracting design tokens rather than guessing from screenshots — rebuilding/editing an interface from code yields far higher quality than from screenshots.

When the Task Involves a Specific Brand — Asset Protocol

Asset > Spec. A brand's identity is "being recognized." Recognition is driven by assets in this order — not by hex codes:

Hard rules:

• Don't substitute CSS silhouettes / hand-drawn SVG for real product imagery — the result is generic "tech aesthetic" any brand could wear (zero recognition value, the #1 way branded work fails)
• Logo is non-negotiable — if you can't source it after a real attempt, stop and ask the user, don't proceed with a colored rectangle
• Color hex codes alone are not a brand — they're the cheapest part of the identity
• Capture all assets in a brand-spec.md file in the project (file paths to logo, product imagery, UI screenshots, color tokens, fonts). All HTML must reference these via <img src="…">, not redraw them

Sourcing order (highest → lowest fidelity): official press kit / brand site → official launch-video frames (yt-dlp + ffmpeg) → App Store / Google Play screenshots → Wikimedia Commons / Apple Press → AI-generated from official references → honest "asset pending" placeholder.

When Adding to an Existing UI

Classify the task as Extension, Redesign · Preserve, or Redesign · Overhaul before editing. Read references/redesign-protocol.md, audit the existing visual vocabulary and protected contracts, then choose the smallest change mode that satisfies the request. New elements in Extension mode should be indistinguishable from the originals.

Step 2b: Produce a Design Read and Calibrate Five Dials

Before choosing tokens, summarize the brief in one concise block. Infer rather than interrogate when context is sufficient:

Design Read:
  artifact: [landing / dashboard / prototype / slides / visualization / ...]
  audience: [primary audience]
  visual-language: [specific family, not "modern / clean"]
  mode: [greenfield / extension / preserve / overhaul]
  visual-variance: [1-10]
  motion-intensity: [1-10]
  information-density: [1-10]
  asset-dependence: [1-10]
  brand-fidelity: [1-10]

Use the dials as decision variables, not decorative scores. They must affect layout variation, motion, content per viewport, real-asset effort, and preservation strictness. Read references/design-calibration.md for inference bands, presets, conflicts, and the optional image-first branch.

Step 3a: Position Four Questions Before Picking a System

Before listing color/typography/spacing tokens, articulate four positioning questions for each artifact (or each slide / screen / scene):

• Narrative role: Hero / transition / data / pull-quote / closing? (Each demands a different visual register.)
• Viewing distance: 10cm phone / 1m laptop / 10m projector? (Drives type scale and information density.)
• Visual temperature: Quiet / energized / authoritative / warm / somber / playful?
• Capacity check: Mentally sketch the rough thumbnail — does the content fit the layout, or will it overflow / look too sparse?

The system that follows must serve these answers. Picking aesthetics in a vacuum is the root cause of generic output.

Step 3: Declare the Design System Before Writing Code

Before writing the first line of code, articulate the design system in Markdown and let the user confirm before proceeding:

Design Decisions:
- Design Read: [one-line synthesis + five dials]
- Anchor / recipe (if any): [e.g., "linear" → `references/style-recipes/linear.md`, or "custom"]
- Color palette: [primary / secondary / neutral / accent]
- Typography: [heading font / body font / code font]
- Spacing system: [base unit and multiples]
- Border-radius strategy: [large / small / sharp]
- Shadow hierarchy: [elevation 1–5]
- Motion style: [easing curves / duration / trigger]

If you picked a recipe from references/style-recipes/, paste its concrete palette / typography / spacing / radius / shadow / motion values straight into the block above — that catalog exists so you don't have to invent these on the fly, which is the leading cause of AI-default Inter + #3b82f6 mush. Load only the one recipe file you're using, not the whole catalog.

🛑 Checkpoint 1: After articulating Steps 3a + 3, stop. Tell the user "I plan to use this system. Confirm and I'll start the v0." Then actually wait — don't say it and immediately start coding.

Step 4: Show a v0 Draft Early

Don't hold back a big reveal. Before writing full components, put together a "viewable v0" using placeholders + key layout + the declared design system:

• The goal of v0: let the user course-correct early — Is the tone right? Is the layout direction right? Are the variant directions right?
• Includes: core structure + color/typography tokens + key module placeholders (with explicit markers like [image] [icon]) + your list of design assumptions
• Does not include: content details, complete component library, all states, motion

A v0 with assumptions and placeholders is more valuable than a "perfect v1" that took 3x the time — if the direction is wrong, the latter has to be scrapped entirely.

🛑 Checkpoint 2: Push v0 to the user before continuing. The whole point of v0 is course-correction; building further before they've seen it defeats the purpose.

Step 5: Full Build

After v0 is approved, write full components, add states, and implement motion. Follow the technical specifications and design principles below.

🛑 Checkpoint 3: When you hit a non-trivial decision point during the build (interaction approach choice, content variant, fundamental layout shift), pause and confirm again — don't silently push through.

Step 6: Verification

Always run the lightweight Pre-delivery Checklist as a code/design self-check.

Run an executable browser acceptance harness only when the user explicitly asks for acceptance / 验收, QA, browser testing / 浏览器测试, visual regression, responsive testing / 响应式检查, cross-viewport verification, or equivalent hands-on validation. Do not infer this request merely from “build,” “finish,” “polish,” or “verify your work.” When triggered, read and follow references/browser-acceptance.md; report evidence and repair failures before delivery.

Step 7: Critique on Request (or as Self-Check Before Delivery)

When the user asks "review this", "is it good?", "score this", "好不好看", or you want to do a self-check before declaring done, run a 5-dimension critique:

Score each 0–10; report overall score, dimension scores, Keep, severity-sorted Fixes, and three Quick Wins. Critique the design, not the designer. Read references/critique-guide.md for the exact format, weighting, issue catalog, and detailed rubrics.

Fallback: Design Direction Advisor

When to trigger:

• The request is genuinely ambiguous ("make something nice", "I don't know what style I want", "give me some directions")
• No design context exists, and the user can't or won't provide reference material
• The user explicitly asks "recommend a style" / "give me a few directions" / "pick a vibe"

When to skip:

• The user already provided a Figma / screenshots / brand reference → go straight to the main workflow
• The user stated a specific direction ("make an Apple-Silicon-style launch animation") → main workflow
• Small tweaks or explicit tool calls ("convert this HTML to PDF") → skip

Mechanism: 3 differentiated directions, not 10 questions

Don't ask the user 10 generic taste questions. Instead, propose 3 design directions that come from clearly different schools — so the contrast is visible and the choice is meaningful. Each direction must include:

• A named designer or studio reference (e.g., "Pentagram-style information architecture", not just "minimalist")
• 2–3 lines of why this direction fits the user's context
• Signature visual cues (3–4 concrete details: color, typography, layout, motion)
• Optional: one famous touchstone work

School library — pick 3 from different rows

❌ Hard rule: never recommend 3 picks from the same row — the user can't tell them apart and the contrast that makes the choice meaningful collapses.

After the user picks

The chosen direction becomes the design context for Step 2 onward. Document it in brand-spec.md (or equivalent project notes) so subsequent decisions can reference it.

Direction → concrete starting point: once the user picks a school, surface 2–3 named recipes from that school by reading the matching files in references/style-recipes/ (e.g., picked Information Architecture → read references/style-recipes/pentagram.md, references/style-recipes/bloomberg-terminal.md, etc.). Each recipe file brings concrete palette, typography, spacing, and signature moves you can paste into the Step 3 design-system declaration.

Extended philosophy library, per-school anchor tables, and AI-prompt templates → references/design-directions.md. Anchored recipe catalog → references/style-recipes/INDEX.md (catalog index + 3 indexes + cross-cutting anti-patterns) + 25 single-recipe files alongside it.

Technical Specifications

React + Babel (Inline JSX)

For React prototypes, use pinned-version CDN scripts with integrity hashes — see the exact <script> tags in references/advanced-patterns.md. Do not change versions, do not add type="module" (breaks the Babel transpilation pipeline). Import order: React → ReactDOM → Babel → your component files.

Three Non-negotiable Hard Rules

1. Never use const styles = { ... } — multiple component files with styles as a global object will silently overwrite each other. Always namespace: const terminalStyles = { ... }, const headerStyles = { ... }. Or use inline style={{...}} directly. Never use styles as a variable name.

2. Separate <script type="text/babel"> blocks do not share scope — each Babel script is compiled independently. To share components across files, explicitly attach them to window at the end of each file: Object.assign(window, { Terminal, Line });

3. Do not use scrollIntoView — in iframe-embedded preview environments, it disrupts outer-frame scrolling. Use element.scrollTop = ... or window.scrollTo({...}) instead.

CSS Best Practices

• Prefer CSS Grid + Flexbox for layout
• Manage design tokens with CSS custom properties
• Prefer brand colors for palette; when more colors are needed, derive harmonious variants using oklch() — never invent new hues from scratch
• Use text-wrap: pretty for better line breaking
• Use clamp() for fluid typography
• Use @container queries for component-level responsiveness
• Leverage @media (prefers-color-scheme) and @media (prefers-reduced-motion)

File Management

• Use descriptive filenames: Landing Page.html, Dashboard Prototype.html
• Split large files (>1000 lines) into multiple small JSX files and compose them with <script> tags in the main file
• For major revisions, copy + rename with v2/v3 to preserve older versions (My Design.html → My Design v2.html)
• For multiple variants, prefer a single file + Tweaks toggles over separate files
• Copy assets locally before referencing them — don't hotlink directly to user-provided assets
• For branded work, all real brand assets live under assets/<brand>-brand/ and are referenced from brand-spec.md

📚 More code templates (device frames, slide engine, animation timeline, Tweaks panel, dark mode, design canvas, data visualization) → references/advanced-patterns.md

Design Principles

Avoid AI-Style Clichés (the WHY matters)

Anti-cliché is not aesthetic snobbery — it's protecting the user's brand recognition. The reasoning chain:

1. The user wants their brand to be recognized
2. AI defaults = average of training data = all brands averaged together = no brand recognized
3. So AI-default output dilutes the user's identity into "yet another AI-generated page"

This is why the only legitimate exception to every anti-cliché rule below is "the brand spec uses it" — at that point it stops being slop and becomes a brand signature.

These are baseline examples, not the whole taxonomy. When designing a multi-section marketing page, redesign, dashboard, or motion-heavy artifact, read only the matching parts of references/failure-patterns.md. Treat each pattern as default → reason → exceptions → detection → repair, not as an unconditional aesthetic ban.