Figma Codegen

figma-codegen

Turn a Figma selection into code that looks like the rest of the project: reuse the components and tokens that already exist, only build what's genuinely missing. The grounded tools do the heavy lifting so you are not guessing from a screenshot. This file is the router; deep detail lives in `references/` — load a reference when its step is in play.

When to use

• The user pastes a Figma URL/selection and asks for code ("code this", "build this component").
• The user wants to extend an existing component to match a Figma frame.

More than one Figma file open

Every tool result says so when it applies, so you do not have to check up front: with two or more files connected and none claimed, results carry a MORE THAN ONE FIGMA FILE IS OPEN block naming them. It matters because calls otherwise follow whichever file the user last touched, which changes when they switch tabs — so a task can read one file and write to another and report success.

When you see it, claim the file before doing anything else: list_files names every connected file, use_file({ fileName }) claims one for this session (use_file({ sessionId }) when two open files share a name — Figma allows that, and x (Copy) is how it happens). Every later call then reaches that file whatever is in front, including while its tab sits in the background. If the user has not said which file, ask — do not guess from the names.

Workflow

Run the grounded tools against the selection, then generate — trust them over the rendered image.

1. `get_design_context` (detail full, dedupeComponents: true) → the structural tree with tokens resolved to names (Primary/500, spacing/4), styles deduped into globalVars, and each instance's mainComponent / componentProperties. This is the layout + binding source of truth. Keep dedupeComponents: true and don't depth-limit a subtree you'll build from. → How to read the tree, the per-node fidelity catalog (effects, per-side borders, stroke align, per-corner radius, blend, masks, gradients, image-fit, auto-layout/grid, aspect-ratio, scroll/sticky, stack & stroke-space), Dev Mode annotations as ground truth, and grounding a page too big for one call: [`references/grounding.md`](./references/grounding.md).

1. `component_map` → every Figma component grouped to a local code component with a status (high / medium / low / unmapped), candidate.filePath, and matchedProps.

• high / medium: reuse that component (import from candidate.filePath), don't regenerate. Never invent a component name component_map didn't report.
• candidate.ambiguousWith (a capped list of { name, filePath }): the Figma name matched two or more code components nearly equally and the join couldn't confidently pick — a verify-me pick, not a confident reuse (the analogue of token_map's ambiguousWith). Check which of the winning candidate + these runner-ups is the right component for _this_ context before importing (a wrong reuse is a silent visual bug), then record the confirmed one in the map file so the next run is certain. Absent when the pick was unambiguous.
• Wire each entry's instances[].props (resolved variant / boolean / text values) onto the reused component — one element per instance, with its own props.
• candidate.unmatchedProps: Figma axes the component has no prop for (a leading icon, a required flag, an active state) → surface as component-extension TODOs, never fake them with ad-hoc markup.
• unmapped: build it new in the project's style. For a repeated unmapped component (instanceCount > 1), build from its first instance's subtree; if that came back deduped/truncated, drill get_design_context on instances[0].nodeId once — don't rebuild a repeated component by eye.
• When you're emitting a component's own definition (its prop types, not just rendering it), get_component_api on the component/instance returns the full property API — every VARIANT option and each BOOLEAN/TEXT/INSTANCE_SWAP prop with its default — so the prop space is grounded, not inferred from the instances you happened to see.

1. `token_map` → every Figma variable joined to a project token with status + ref + matchedBy. A document's shared paint styles (single solid color styles — the token mechanism of pre-variables files) join the same way, marked source: 'style'.

• mapped: reference candidate.ref (bg-primary-500, var(--color-primary-500)) — never the raw hex/px get_design_context resolved. matchedBy: ['name'] on a colour (value drifted): use it but flag the mismatch.
• candidate.ambiguousWith (several project tokens share that exact value and the name couldn't split them): a capped, verify-me pick — choose the semantically right sibling for the context (or keep the value and flag the gap), never treat it as a confirmed reuse.
• On a document with few or no variables (most real-world files), get_design_context's own projectTokens map is the fallback: any raw color in the payload that exactly equals a project token's value is annotated there ({ "#6266F0": { ref, name, matchedBy: ["value"] } }). Before hardcoding a hex, look it up and emit the ref when it fits the context semantically. matchedBy: ["value"] marks the entry as name-blind value-equality evidence — a hypothesis to verify, not a resolved binding: an entry with candidates lists same-value tokens to choose between by meaning, a semantically wrong token is worse than the raw value (keep the raw value and note the gap instead), and a bound Figma variable always outranks a raw-value match.
• framework-builtin (a built-in scale step of the project's utility framework — Tailwind or UnoCSS — e.g. spacing/4, line-height/7, weight/Bold): carries builtin: { scale, step } — compose the utility (p-4/gap-4, leading-7, font-bold), not an arbitrary p-[16px]. This is not a gap.
• from (a SCSS variable): the ref does not resolve on its own. The file you write must import the declaring file, and from is repo-relative while Sass resolves @use against the _importing_ file — so re-resolve it from where you are writing: from src/components/card.scss, from: src/styles/_tokens.scss becomes @use '../styles/tokens' as *, never the repo-relative path verbatim. as * keeps the ref as given; the project's own @use style may namespace it instead (@use '../styles/tokens' makes $color-primary-500 into tokens.$color-primary-500). Emitting the ref without the import is a compile error, not a style nit.
• unmapped: use the value but call out the gap (offer to add it to the token source); don't hardcode silently.
• figmaModes ({ Light: …, Dark: … }, with the file's theme axes on themedCollections): the token is theme-dependent — figmaValue is only the default mode, never the whole story. Mapped: emit candidate.ref and confirm the project token itself switches per theme (.dark / [data-theme] / prefers-color-scheme / a dark: story); if the project defines only one value, wire the other mode's value into that mechanism or surface a theme gap. Unmapped: the default mode is the base, other modes ride the project's dark-mode mechanism (dark: variants / prefers-color-scheme). Themes encoded without native modes — paired collections or name groups (Color/Light/* + Color/Dark/*, a plan-limited workaround) — get the same treatment.

1. Export the assets grounding can't carry — logos, photos, icons have no pixels and otherwise render as grey blocks. save_image_fills for IMAGE-fill nodes (the original asset, not a