Figma Extract Design System

figma-extract-design-system — production app(s) → design system

Organizations let small teams vibe-code products with AI. The products work — but there's no design system underneath, so nothing is consistent, reusable, or versionable. This skill runs the extraction engagement: app(s) → component inventory → DTCG tokens → Storybook package → (optionally) Figma.

Two valid stopping points:

• Code-led org: stop at the Storybook package. It becomes the versioned source of truth that gets packaged and re-imported into the products.
• Design-led org: continue into Figma — tokens become variables (figma_import_tokens), key components become Figma component sets.

Tooling modes

• Preferred — figma-console-mcp connected (Local Mode): use the five figma_ds_* tools. They are deterministic, bounded, and persist manifests under <outDir>/.extraction/ so the engagement survives session boundaries.
• Fallback — no console MCP: do the same phases yourself with Read/Grep/Bash (slower, less deterministic). Keep the same artifacts: an analysis.json you write yourself, DTCG tokens/tokens.json, and the same package layout. For the Figma round-trip use the figma-import-tokens skill (official Figma MCP use_figma).

Phase 0 — Discovery (talk first)

Establish with the user before scanning:

1. Target app path(s) — absolute paths. Multiple apps is a feature: the real design system is the intersection of their products.
2. Code-led or design-led? Decides whether Phase 6 (Figma) happens.
3. Brand/naming context — product name, existing brand color names if any. Needed for the semantic-naming pass.
4. Where the design system lives — default <first app>/design-system, or a separate dir in the monorepo.

Phase 1 — Analyze

Copy code
figma_ds_analyze { targets: ["/abs/path/app"], outDir: "/abs/path/app/design-system" }

Review the summary with the user before proceeding — especially the architecture block, which is what separates a design system from a UI kit:

• specializations are usage-named components (FollowButton, ModerationMenu, NodeCard) — each is really a generic component (Button, Menu, Card) plus a usage. The usage is what makes it a "follow button"; the component is just a button.
• missingPrimitives is the build list: generics the app implies but never made.
• levels is the atomic-design read (atom/molecule/organism/template) — heuristic; confirm the org's vernacular with the user.

Generalization pass (after review, before/during porting): if the design-systems-assistant MCP is connected, ground decisions in it (search_design_knowledge for naming conventions, component API design, atomic hierarchy — it carries the cross-industry best practices). Then:

1. Build each agreed missingPrimitive as a generic component with a variant/prop API wide enough to cover its observed specializations.
2. Re-express each specialization as a recipe: either a story on the generic ("Button → Recipes/Follow") or a thin wrapper in a recipes/ tier — never a sibling primitive. The product keeps its follow button; the SYSTEM ships Button.
3. Organize Storybook by the chosen layers (e.g. Primitives/, Components/, Patterns/) instead of one flat Components/ list.
4. Not every specialization collapses — positional qualifiers (LeftNav) or true one-off organisms (ThreadCanvas) may stay; the report proposes, the user disposes.

Also review the classic items:

• Classification matters. vendored (shadcn-style copied source) and bespoke get ported as-is; wrapped gets ported keeping the vendor package as a peer dep; pure-vendor (raw MUI/Chakra usage) is NOT ported — it's represented by theme/tokens and usage docs.
• duplicateGroups are the vibe-code smell (three Button.tsx files). Decide the canonical one per group with the user; record losers as skipped in figma_ds_status with a note.
• Porting order = the tool's usage ranking (topByUsage). Don't port long-tail components nobody uses without asking.

Phase 2 — Extract tokens

Copy code
figma_ds_extract_tokens { outDir, formats: [...match the app's styling method...] }

• declared tokens (CSS vars, @theme, SCSS vars, tailwind theme, shadcn HSL triples) keep their names and carry sources provenance. .dark / [data-theme=dark] values become a real Light/Dark multi-mode structure.
• inferred tokens are frequency-promoted raw values — review them with the user; the belowThreshold list is worth a skim for near-misses.

Semantic-naming pass (your job, not the tool's): structural names (color/blue/500, color/extracted/hex-22c55e) are correct but not shippable. With the user's brand context, layer semantic alias tokens (color/primary → {color.blue.500}) by editing tokens/tokens.json — add to the Semantic group, reference primitives with {dot.path} aliases, never delete the primitives. Re-run figma_ds_extract_tokens only if the codebase changed (it regenerates; hand edits to generated projections are lost, edits to tokens.json are the input for scaffold).

Phase 3 — Scaffold + Storybook

Copy code
figma_ds_scaffold { outDir, packageName: "@org/design-system" }
cd <outDir> && npm install && npm create storybook@latest -- --yes --no-dev
figma_ds_setup_storybook { outDir }    # AFTER init — wires the workshop to the extraction

figma_ds_setup_storybook generates .storybook/preview.css (extracted tokens + the source app's @theme mapping, @utility definitions, @layer base, and @font-face rules so production class names and fonts render faithfully), copies self-hosted fonts, and patches main.js (tailwind vite plugin + automatic JSX runtime) and preview.jsx (theme toolbar/decorator from the extracted mode names). Apply any manualSteps it returns, then restart the dev server. Verify .storybook/main story globs include src/**/*.mdx — the generated src/docs/*.mdx pages are the token/typography/iconography showcases.

Phase 4 — Port components (the loop)

Per component, in usage-rank order:

1. figma_ds_extract_component { outDir, component: "Button" } — gives source, local import closure, prop contract, observed call-site variants, style touchpoints, and a CSF3 story scaffold.
2. Port into <outDir>/src/components/<Name>/ — each component gets its own directory, component file, stories file, and index (never monolithic files). Decouple app-specific imports (routing, data fetching, global state); replace hardcoded values with tokens (check styleTouchpoints.customProperties against tokens.json).
3. Adapt the story scaffold — one story per real variant (observedProps), tags: ['autodocs'].
4. Visually verify in Storybook (browser) before calling it done.
5. figma_ds_status { outDir, update: { component, status: "ported", storyFile } }.

Long engagements: figma_ds_status { outDir } at session start tells you exactly where you left off.

Phase 5 — Code-led stopping point

Buildable, versioned package: components + tokens + Storybook. Set up whatever publishing/versioning the org uses (changesets, private registry). The products then consume @org/design-system instead of their local copies — that migration is a separate engagement.

Phase 6 — Design-led: into Figma

1. Tokens → variables: figma_import_tokens with tokens/tokens.json (top-level groups become collections; Light/Dark modes come across). Preview with strategy: "dry-run" first.
2. Components → Figma: rebuild key components as component sets using the existing tools (figma_create_component_set, figma_execute) or the figma-generate-library / prototype-to-figma workflows — bind fills/strokes/radii to the imported variables so Figma and code share one token source.
3. From then on the sync loop is figma_export_tokens / figma_import_tokens diff-aware merge.

Boundaries & gotchas

• figma_ds_* tools are Local Mode only (they read/write the local filesystem).
• Extraction never mutates the target app — everything lands under outDir.
• Absolute paths always: the MCP server's working directory is not the project.
• Don't hand-edit generated projections (tokens.css, theme.css, …); edit tokens.json (or the source app) and regenerate.
• Provenance lives in $extensions["figma-console-mcp.extraction"] on every token — use it to justify decisions to the user ("this is your primary because it appears 214 times").