Agentic Design Systems

Agentic Design Systems

When this skill triggers, you're helping the user build, extend, or audit a component library whose primary consumer is an AI agent. The bar: given a prose request like "Build a confirmation modal with a destructive action," an agent picks the right component, the right variant, and the right tokens — without inventing patterns.

This skill gives you the schema, the workflow, and the principles. Adapt it to where the user is.

---

Diagnose first

Before producing anything, figure out which mode applies:

• Greenfield — user is starting a new system intended for agent consumption. Walk through the build workflow. Start with the schema and one worked component (Button is canonical); don't try to scaffold everything at once.
• Retrofitting — user has a component library and wants to make it agent-readable. Skip workspace setup. Add meta.types.ts, then write a .meta.ts per component, then build the index and validator. Anti-patterns first (see Step 5).
• Auditing — user has metadata already. Score each component against the four pillars and the validator checks. Flag missing relationships, prose anti-patterns, raw global tokens, and ungrounded variant axes.
• Single component — user wants to add or fix one component. Generate the full file set in Step 3. Metadata ships with the component or it doesn't ship.

If the prompt is ambiguous, ask one targeted question — don't guess.

---

The four pillars

Treat every component as the intersection of four things. Most metadata schemas only model the first; that's why agents misuse the components.

Plus aiHints — the meta layer that tells an agent when to use the component, which variant fits which situation, and what never to do.

Four design decisions baked into the schema

• States are implicit in tokens, not a separate pillar. Encode button-primary-bg-hover, button-primary-bg-disabled, etc. Theming becomes a token swap; one source of truth.
• Variants are a matrix, not a flat enum. Declare axes (appearance × size × density) and let the agent pick a cell. Use invalidCombinations for cells that shouldn't ship.
• Accessibility folds into Relationships. ARIA role, keyboardSupport, and screenReader describe how the component fits into the document and interaction model — that's relational. No separate a11y pillar.
• Anti-patterns are first-class and structured. Never prose. Always {scenario, reason, alternative} triples — they force precision.

---

The schema (canonical contract)

Write this once at meta.types.ts. Every component's .meta.ts imports it and is type-checked. The schema is the contract.

interface ComponentMeta {
  component: {
    name: string;
    category: "atoms" | "molecules" | "organisms";
    type: "interactive" | "display" | "container" | "input" | "navigation";
    description: string;
    path: string;
    figma?: { nodeId: string | null };
  };

  props: Record<string, PropDef>;

  variants: {
    axes: Record<string, readonly string[]>;
    purpose: Record<`${string}.${string}`, string>;
    invalidCombinations?: { axes: Record<string, string>; reason: string }[];
  };

  relationships: {
    requires?: string[];           // contexts/providers above
    mustBeChildOf?: string[];
    mustBeParentOf?: string[];
    optionalSibling?: string[];
    commonPartners?: string[];
    triggers?: string[];           // events emitted
    blocksWhen?: { when: string; effect: string }[];
    exposesState?: string[];       // state descendants can read
    role: string;                  // a11y
    keyboardSupport: string;
    screenReader: string;
  };

  tokens: {
    color?: Record<string, string>;
    spacing?: Record<string, string>;
    typography?: Record<string, string>;
    border?: Record<string, string>;
    motion?: Record<string, string>;
    elevation?: Record<string, string>;
  };

  aiHints: {
    priority: "high" | "medium" | "low";
    keywords: string[];
    selectionCriteria: Record<string, string>;
    usage: {
      useCases: string[];
      commonPatterns: { name: string; composition: string }[];
      antiPatterns: { scenario: string; reason: string; alternative: string }[];
    };
  };
}

When generating a component's .meta.ts, fill every field that applies. Empty arrays are fine; missing keys aren't — that's what the validator catches.

---

Build workflow

Step 1 — Workspace shape

Recommend a sibling package inside the consuming app's monorepo (packages/ui-next/). Switchover later is just an import rewrite. A separate repo only makes sense when independent versioning is required.

Step 2 — Schema as TypeScript contract

Write meta.types.ts before any components. Everything else flows from it.

Step 3 — Build one component end-to-end

Pick something small and high-traffic. Button is canonical. Ship the full set together:

Button/
  Button.tsx              ← implementation
  Button.meta.ts          ← four pillars + aiHints
  Button.tokens.css       ← component-scoped tokens
  Button.stories.tsx      ← visual test surface
  Button.test.tsx         ← behavior tests
  index.ts                ← single canonical export

Wire up Storybook early (.storybook/main.ts for story discovery, .storybook/preview.ts for theme imports). One story per variant matrix cell makes design-space regressions obvious at a glance.

Metadata ships with the component or the component doesn't ship. Don't defer it.

Step 4 — Tokens at two levels

• tokens/core.css — raw brand palette and scales (--color-brand-600, --space-4).
• tokens/themes/*.css — map component-scoped tokens (--button-primary-bg) to core values.

Component CSS only ever references component-scoped tokens. Theme swaps stay mechanical.

Step 5 — Anti-patterns first

For each component, write the antiPatterns array before writing the implementation. The structured-triple format (scenario, reason, alternative) forces precision — you can't write "don't overuse primary buttons"; you have to write which scenario, why it's wrong, what to do instead. The anti-patterns end up driving the API.

Step 6 — Variant axes as a coordinate system

Variants aren't a flat list of strings. Declare axes (appearance × size × density) so the agent picks along independent dimensions. Use invalidCombinations to rule out cells that shouldn't ship (e.g. appearance: "ghost" × size: "xs" — too small to be tappable).

Step 7 — Relationships as machine-checkable rules

• requires — providers that must exist above
• mustBeChildOf / mustBeParentOf — structural constraints
• triggers — events emitted
• blocksWhen — prop-state-dependent behavior
• exposesState — what descendants can read

This is the pillar that prevents agents from generating code that compiles but is structurally wrong.

Step 8 — Hierarchical metadata index

Generate metadata/index.json — a flat list of {name, category, path, keywords, priority}. Agents scan this first to shortlist candidates, then read the full .meta.ts only for relevant components. Cheap discovery, expensive depth.

Build with a script (scripts/build-index.ts) that walks every *.meta.ts, dynamically imports each one, and writes the JSON. Run it on every metadata change so the index can never lie about what exists.

Step 9 — Metadata validator

A short script (scripts/validate-metadata.ts) walks every *.meta.ts (via fast-glob), dynamically imports each one, shape-checks the export, and asserts:

• Every variant axis cell appears in aiHints.selectionCriteria or variants.purpose
• Every tokens.* key is component-scoped (kebab-case of the component name)
• antiPatterns is non-empty for priority: "high" components
• relationships.role / keyboardSupport / screenReader are non-empty (enforces the "a11y folds into Relationships" decision)
• invalidCombinations references only declared axis values

Run in CI. If the metadata is wrong, the build fails. The schema is enforced, not aspirational.

Step 10 — Switchover (per-component, not big-bang)

Don't switch over at the end — switch per-component as parity is reached.

• API parity tracker — markdown checklist of old library exports → new equivalents.
• Token bridge — map old tokens to new component-scoped ones.
• Codemod — rewrite imports page-by-page.

---

Principles

• Lean over generated. Hand-write metadata for the first ~10 components. Automate only after the patterns are obvious.
• Direct imports, no barrels. One canonical path per component. Two ways in means the agent picks wrong.
• Co-locate everything. Component, metadata, tokens, stories, tests — one folder.
• Anti-patterns drive design. Write them first. They reveal the contract.
• Component-scoped tokens. Never reference raw global tokens from a component's CSS.
• States live in tokens. No states block. No behavior.states array.
• A11y is relational. It belongs in Relationships, not its own pillar.
• The schema is the contract. Validator runs in CI; failing metadata fails the build.

---

What success looks like

The user hands an agent a Figma screenshot or a prose request — "Build a confirmation modal with a destructive action" — and the agent:

1. Scans metadata/index.json and shortlists Modal, Button, Heading. 2. Reads each shortlisted component's .meta.ts. 3. Sees Button.relationships.mustBeChildOf includes ModalFooter; picks appearance: "danger" from aiHints.selectionCriteria. 4. Avoids two appearance: "primary" siblings because antiPatterns flags it. 5. References component-scoped tokens — never invents a color. 6. Generates code that follows the contract on the first try.

That's the bar. If the metadata can't get an agent to that outcome, fix the metadata.