Design System

Build a systematic design system

frontend-design decides how the product should look; this pack decides how that look is structured so it scales. A design system is the disciplined layer underneath the aesthetic: a three-tier token architecture, component specs that cover every interactive state, and a deliberate match between the product's needs and the system it gets. Get this right and every later surface reuses tokens instead of re-inventing styling per component.

Keep SKILL.md lean — the depth lives in the references; load the one you need on demand:

The token architecture in one breath

Three layers, each referencing the one below — never skip a layer:

/* PRIMITIVE — raw values, no meaning. Change rarely. */
--blue-600: oklch(55% 0.20 264);
/* SEMANTIC — purpose aliases. This is the theme-switch seam. */
--color-primary: var(--blue-600);
/* COMPONENT — per-component knobs. Change freely. */
--button-bg: var(--color-primary);

Why it matters: theming happens at the semantic layer (override --color-primary, every component follows). Per-component tweaks happen at the component layer without disturbing anyone else. Components reference component or semantic tokens — never a primitive directly, and never a raw hex/px. See references/token-architecture.md.

House idiom: Gaffer uses oklch for colour and clamp() for fluid type at the primitive layer (perceptual uniformity, predictable contrast). Adapt the references' hex examples to oklch to stay consistent with brand and frontend-design.

Recommend a tailored design system

When the ask is open-ended ("what should this product use?"), don't reach for a generic kit. Reason from the product to a specific system across four axes:

1. Pattern / structure — content-first marketing site, data-dense dashboard/admin, a transactional flow, or a component-library/SaaS surface. This sets density, the spacing rhythm, and how much the grid is allowed to break.
2. Style direction — defer to brand if one exists; otherwise pick a real one (editorial, Swiss, neo-brutalism, glassmorphism-with-depth, light/dark luxury, bento) and say why it fits this product. Never "clean minimal".
3. Palette posture — how many primitives, which semantic roles (primary + secondary + muted + the four status colours), and whether dark mode is a real requirement or a habit. Use the 60/30/10 dominant/secondary/accent ratio as a sanity check, not a law.
4. A11y + performance posture — contrast floors (4.5:1 text, 3:1 large/UI), visible focus, prefers-reduced-motion, plus the CWV/bundle budgets from frontend-design. State these as targets up front so they constrain the system, not get bolted on after.

Output the recommendation as a short rationale + the token + component scaffold it implies — then hand off to frontend-design for the visual execution and brand for the identity.

Component spec discipline

A component isn't "done" until every state is specified, not just the default. For each component define variants (default/secondary/outline/ghost/destructive…), sizes (sm/default/lg with explicit height + padding + font), and the full state matrix:

State priority when several apply: disabled > loading > active > focus > hover > default. The full matrices for button/input/card/badge/alert/dialog/table live in references/component-specs.md; the cross-component state + variant rules in references/states-and-variants.md.

Steps

1. Read the lore + existing tokens first. search_lore (Memory MCP) for the design system and any token ADRs; inspect tokens.css/theme files and a sibling component. If a token system or brand already exists, extend it in its idiom — do not introduce a competing one.
2. Establish the three layers (or audit the existing set against them). Primitives as raw oklch/clamp() values; semantic aliases for every role; component tokens per component. Add a .dark block that overrides semantic tokens only.
3. Refuse primitive/hardcoded leaks in components. Components consume semantic/component tokens exclusively — no raw hex, no px font sizes, no primitive references.
4. Specify components fully — variants, sizes, and the complete state matrix above, with accessible focus and ARIA states. An unspecified state is a bug waiting to happen.
5. If asked to recommend a system, run the four-axis reasoning above and output a rationale + scaffold; hand visual execution to frontend-design, identity to brand.
6. Verify + evidence. Run the repo's tests + lint; for token work, grep components for raw hex / px font-size leaks. Record test_output via record-evidence and submit for review — never self-approve.

Build / Test

• Run the repo's configured test + lint; type-check (tsc --noEmit) for TS surfaces.
• Token-compliance check: grep component styles for raw hex and px font sizes — they should reference tokens, not literals.
• Verify focus is visible on every interactive component and that .dark only overrides semantic tokens (not primitives or component tokens).
• The DoD is verified by the repo's configured commands — record the output as evidence.

Rules

• Three layers, no skipping: components reference semantic/component tokens, never a primitive or a raw hex/px.
• Theme at the semantic layer; tweak at the component layer; touch primitives rarely.
• Every component spec covers the full state matrix with a visible, accessible focus ring.
• Recommend a specific system from product → pattern + style + palette + a11y/perf — defer to brand for identity and frontend-design for visual execution; never duplicate them.
• oklch colour + clamp() type at the primitive layer to stay consistent with the house packs.

Capture lore

A repo's token architecture — the layer split, the naming convention, the dark-mode seam, or a "never reference a primitive in a component" rule — is exactly the fact the next agent needs before they start. That kind of fact is lore. Capture it via the lore-capture protocol in your brief (CLAUDE.factory.md, step 11 "Memory contribution"): call the Memory MCP suggest_lore once at the close of your work — reusable conventions, gotchas, decisions, and boundaries only, never per-ticket trivia.