Teaching Site Design System

Token starter (copy this, don't re-derive): templates/tokens.css — full :root block + dark theme + signature components (day-hero, glass-card, prompt-card, task-list, learning-goal, concept, unit, material, table). Paste into the <style> block of index.html (vanilla SPA pattern) or use as external style.css.

Schema authority: this skill defines visual tokens; data field names live in _shared/domain-primitives.md. When styling a component, read that file to know the data shape you're rendering.

Reference implementation: d:/GitHub/ai-workshop/index.html:11-1700 — the full production CSS this skill's tokens were extracted from.

This skill defines the visual layer of a teaching site: color tokens, typography, geometry, components, dark mode strategy, and — critically — the reasoning behind each decision. The rationale matters more than the literal values; future sites can re-tune the values, but the why survives across projects.

When to Invoke

• Setting up a new teaching site (read tokens before writing first CSS).
• Adding a new visual element (badge, card, hero) — check whether an existing pattern covers it before inventing.
• Tuning dark mode (which token to override, which not to).
• Producing a corporate edition or ebook (same tokens flow to print CSS).
• Reviewing visual consistency complaints ("it looks inconsistent here").

The Foundational Decision: OKLCH, Not HEX/RGB

:root {
  --bg:      oklch(99% 0.002 240);
  --surface: oklch(100% 0 0);
  --fg:      oklch(18% 0.012 250);
  --accent:  oklch(58% 0.18 255);
  --accent-2: oklch(66% 0.14 65);
}

Why OKLCH (not #1a73e8 or hsl(...)):

1. Perceptually uniform — oklch(60% ...) and oklch(70% ...) look like an equal step apart; HSL doesn't (its 60→70 yellow jumps visually huge but blue barely changes).
2. Predictable dark mode — for each light token at L=N%, the dark equivalent is roughly L=(100-N)% with the same chroma+hue. Mechanical.
3. color-mix(in oklab, ...) works natively — e.g. soft accent tints (color-mix(in oklab, var(--accent) 18%, transparent)) compose cleanly, no manual rgba math.
4. Modern browser support is fine (Chrome 111+, Safari 15.4+, Firefox 113+). For a static teaching site this is acceptable.

If a stakeholder hands you brand colors as hex: convert once to OKLCH and never look back.

Token Architecture (Two-Axis)

Axis 1: Semantic Surface Tokens

--bg              /* page background */
--surface         /* card / panel */
--surface-2       /* nested card / hover state */
--fg              /* primary text */
--muted           /* secondary text */
--muted-2         /* tertiary text / placeholder */
--border          /* default border */
--border-strong   /* emphasis border */

These are structural — they describe role in the layout, not color. Dark mode swaps the values; component CSS doesn't change.

Axis 2: Accent Tokens (Dual Accent Strategy)

--accent          /* primary action color — blue/violet typically */
--accent-soft     /* faint tint, for hover backgrounds, badges */
--accent-deep     /* high-contrast variant, for hover text */

--accent-2        /* secondary accent — orange/amber typically */
--accent-2-soft
--accent-2-deep

Why two accents (not one + grayscale): teaching sites have two equally important attention paths:

• Action path (--accent): "click me", "copy this", "navigate here" — typically cool color (blue).
• Content path (--accent-2): "look here, this is rich content" — Day hero numerals, PROMPT badges, callout backgrounds. Typically warm color (orange/amber).

Single-accent designs end up using grayscale for content emphasis, which makes everything feel like a button. Dual-accent separates "things you click" from "things you read carefully".

Status Tokens

--success, --success-soft   /* learning goals ✓, copy success */
--warning, --warning-soft   /* gotchas, optional warnings */
--danger,  --danger-soft    /* errors, destructive actions */

Always provide a -soft variant for backgrounds; the base is for text/icons.

Typography: One Font Stack to Rule Them All

--font-display: "Nunito", "jf-openhuninn-1.1", "jf-openhuninn",
                "Gen Jyuu Gothic", "Taipei Sans TC Beta",
                "PingFang TC", "Noto Sans TC",
                system-ui, sans-serif;

/* All four roles point to the same stack */
--font-body:  var(--font-display);
--font-serif: var(--font-display);
--font-mono:  var(--font-display);

Why one stack for everything on a Chinese-language teaching site:

1. Latin + 中文 混排在 90%+ 的段落都會出現。每換一種字體就會出現「字體斷裂感」。
2. The fallback chain handles bilingual rendering — Nunito for Latin glyphs, openhuninn for 中文.
3. Mono / serif variants are mostly used for spacing / weight effects, not actually different typefaces. Implement via font-feature-settings, letter-spacing, font-weight instead of swapping fonts.

Exception: ebook PDF for formal print may use a true serif (Source Han Serif) — that's the only place to break this rule.

Geometry Tokens

--r-sm: 6px;   /* buttons, small chips */
--r-md: 10px;  /* cards, panels */
--r-lg: 14px;  /* heroes, large containers */

--sidebar-w: 268px;
--header-h: 56px;
--content-max: 880px;     /* reading line length cap */

--shadow-lift:
  0 6px 18px -10px oklch(20% 0.02 250 / 0.18),
  0 2px 4px  -2px oklch(20% 0.02 250 / 0.06);

Three radii is enough — more produces visual chaos. Match radius to element scale: button → sm, card → md, hero → lg.

--content-max: 880px is deliberate — comfortable reading line length is ~60–75 CJK characters. Wider feels like documentation, narrower feels like mobile-stuck-on-desktop.

Double-layered shadows (--shadow-lift): one wide-soft + one tight-tighter. Single-layer shadows look digital; double-layer mimics physical card lift. Always negate the y-offset (-10px, -2px) so the shadow doesn't bleed below the element.

Signature Components

1. Day Hero Numeral

.day-hero-numeral {
  font-family: var(--font-serif);
  font-size: clamp(120px, 16vw, 200px);
  font-weight: 500;
  line-height: 0.85;
  letter-spacing: -0.05em;
  color: var(--accent-2);
  text-shadow: 0 1px 0 color-mix(in oklab, var(--accent-2) 18%, transparent);
}
.day-hero:hover .day-hero-numeral {
  color: var(--accent-deep);
  text-shadow:
    0 2px 0 color-mix(in oklab, var(--accent-2) 22%, transparent),
    0 18px 40px color-mix(in oklab, var(--accent) 22%, transparent);
}

Why a giant numeral instead of decorative SVG / illustration for Day headers: an abstract SVG looks like a placeholder ("they forgot to put a real image"). A 200px "D2" reads as intentional, confident, navigational. Specific > abstract.

Why clamp(120px, 16vw, 200px): scales with viewport width but with floor and ceiling so it never feels broken on mobile (would be 120px) or absurd on 4K (capped at 200px).

Why @keyframes fade-in, not opacity: 0 → transition: if IntersectionObserver fails to fire (e.g. CSS zoom is in play, see static-spa-interactions), the numeral stays at opacity: 0 forever. With @keyframes, the default state is visible; animation only "plays once if triggered".

2. PROMPT Block (Orange Badge + Dark Code)

.prompt-block {
  background: oklch(15% 0.02 250);          /* always dark, even in light mode */
  color: oklch(92% 0.005 250);
  border-left: 4px solid var(--accent-2);
  padding: 14px 16px;
  border-radius: var(--r-md);
}
.prompt-block::before {
  content: 'PROMPT';
  display: inline-block;
  background: var(--accent-2);
  color: white;
  padding: 2px 10px;
  border-radius: 999px;
  font-size: 11px;
  font-family: var(--font-mono);
  letter-spacing: 0.1em;
  margin-bottom: 8px;
}

Why orange (--accent-2) for PROMPT: blue (--accent) is reserved for "clickable / navigation". Orange marks "this is a recipe — read carefully and copy verbatim". Three months in, learners associate orange = prompt block automatically.

Why dark background even in light mode: code-y content reads as code-y. Switching light/dark would feel like the prompt is "less serious" in light mode.

3. Glass Card (glass-card)

.glass-card {
  background: color-mix(in oklab, var(--surface) 60%, transparent);
  backdrop-filter: blur(12px) saturate(140%);
  border: 1px solid color-mix(in oklab, var(--border) 80%, transparent);
  border-radius: var(--r-md);
  padding: 16px 20px;
  box-shadow: var(--shadow-lift);
}

Why translucent + backdrop-filter: on a page with rich background (gradient, dotted pattern, illustrations), opaque cards feel heavy. Glass cards let content breathe, signal "this floats on top of the document".

Caveat: requires backdrop-filter browser support. Safari needs -webkit-backdrop-filter prefix. On unsupported browsers, the fallback is just the 60%-opacity color — degrades gracefully.

4. Task Checkbox (Three-State Visual)

.task-checkbox {
  width: 18px; height: 18px;
  border: 1.5px solid var(--border-strong);
  border-radius: var(--r-sm);
  display: grid; place-items: center;
  transition: all 120ms ease;
}
.task-item:hover .task-checkbox { border-color: var(--accent); }
.task-item.done .task-checkbox {
  background: var(--success);
  border-color: var(--success);
}
.task-item.done .task-label { text-decoration: line-through; color: var(--muted); }

Three states: default (empty + strong border), hover (accent border, hint of intent), done (filled green + strikethrough). The hover state matters — without it, learners aren't sure the checkbox is clickable.

5. Goal Badge (Green ✓)

.learning-goal::before {
  content: '✓';
  display: inline-block;
  width: 20px; height: 20px;
  background: var(--success-soft);
  color: var(--success);
  border-radius: 999px;
  text-align: center;
  line-height: 20px;
  margin-right: 8px;
  font-weight: bold;
}

Use the green-on-light-green disc consistently for "learning outcomes" / "you'll be able to ...". Don't reuse for "completed" — that's task-checkbox's job. Green has two distinct semantic uses which must look different: goal (badge prefix) vs done (filled checkbox).

Dark Mode Strategy

:root { /* light tokens here */ }

[data-theme="dark"] {
  --bg:      oklch(16% 0.014 250);
  --surface: oklch(19.5% 0.014 250);
  --fg:      oklch(96% 0.005 250);
  --border:  oklch(28% 0.014 250);
  /* accents usually stay the same — accent-deep may flip to brighter */
  --accent-deep: oklch(72% 0.18 255);
}

Use [data-theme] attribute, not prefers-color-scheme: a manual toggle is the source of truth. Auto-following OS theme means the user can't decide per-site. The toggle reads/writes localStorage.

What flips in dark mode:

• ✅ All surface tokens (bg / surface / fg / muted / border)
• ✅ accent-deep (the high-contrast variant) — flip to a brighter L value
• ❌ Status colors (success / danger / warning) — these are semantic, keep their hue; the -soft variant naturally darkens because of the surface change
• ❌ accent-2 — the dual-accent identity should survive dark mode (the orange brand stays orange)

What stays the same:

• All geometry tokens (radii, sidebar width, shadow definitions — though shadow opacity may need tuning)
• All component layouts and structure

The rule of thumb: token values change, component CSS does not. If you find yourself writing [data-theme="dark"] .task-checkbox { ... }, the abstraction has leaked — refactor the component to use tokens for the property that's varying.

Cross-Format Token Reuse (Web → Ebook)

The ebook print CSS (course-ebook-publishing) reads the same tokens. Specifically:

/* style-ebook.css (print only) */
@media print {
  :root {
    /* Override only what print needs differently */
    --bg: white;
    --surface: white;
    /* But colors of badges, prompts, headings stay the same */
  }
  .prompt-block { /* same .prompt-block style — looks identical to web */ }
  .learning-goal::before { /* same green disc */ }
}

Why: a learner who sees the web first, then opens the PDF handbook, should feel "same product". If web is orange-PROMPT and PDF is gray-PROMPT, the visual brand is shattered.

The corporate edition (course-corporate-edition) does the same — even with brand-customised assets, the design tokens remain.

Three Decisions That Sound Small But Aren't

Decision 1: color-mix(in oklab, X 18%, transparent) for soft tints

Don't hand-write rgba(...). Use color-mix so tints derive from the source token:

background: color-mix(in oklab, var(--accent-2) 18%, transparent);

When you eventually re-tune --accent-2, every tint follows. With manual rgba, you'd have 50 places to update.

Decision 2: No grid system, just --content-max

Avoid bringing in a 12-column grid framework. For a teaching site with linear reading flow, max-width: var(--content-max); margin: 0 auto is enough. Columns appear locally via grid-template-columns inside specific components (figure grid, tool gallery) — but the page itself is a stream.

Decision 3: Animation budget = "almost none"

Allowed:

• fade-in on scroll (one-time per element)
• Hover micro-interactions (color / shadow shift, 120–200ms)
• Sidebar slide (transform, 200ms)

NOT allowed:

• Page transitions between sections
• Decorative idle animations
• Anything > 400ms

Teaching content is dense; animation is a distraction tax. Spend the attention budget on content, not motion.

Anti-Patterns

• Inventing a new color for "this one section" — pull from existing tokens or refactor to add a properly-named token. Drive-by hex values rot the system.
• Using --accent for content-emphasis (like PROMPT badge) — collapses the dual-accent semantic. Use --accent-2.
• Different font for code blocks — see typography section. Use the same stack with font-weight / letter-spacing for differentiation, or accept the visual cost.
• Manually styling dark mode per component — see dark mode section. If a component needs special dark-mode handling, the component is using a non-tokenised value somewhere.
• Forgetting -soft variants — every accent / status color needs a soft variant. Without it, designers reach for ad-hoc opacity values.

Hand-off

When this skill is consumed by another skill (SPA / interactions / ebook / corporate):

• Reference :root tokens, never raw values.
• If you find yourself writing a literal #... hex or rgb(...), you've broken the system — either it belongs as a new token, or you're recreating an existing one.
• New components should compose existing patterns (glass-card + prompt-block + task-list) before inventing.

The token file (the :root { ... } block) is the single source of truth. Print it on the wall. Re-tune values when the brand evolves, but never widen the token set casually.