figma-build

Turn code or a description into a Figma design that looks like it belongs in the file: reuse the components, variables, and styles that already exist, and only build what's genuinely missing. This is the mirror image of figma-codegen — same "reuse beats regenerate / reference tokens not literals" philosophy, pointed at the canvas instead of the codebase. This file is the router; deep detail lives in references/ — load a reference when its step is in play.

You operate on the connected Figma file (the plugin's current session). There is no fetch-by-URL; the user must have the target file open. Confirm a plugin is connected (ping) before building.

When to use

• The user wants to create or update something in Figma from code or a description — a screen, view, modal/dialog/drawer/sidebar/panel, a single component, or a design-system asset.
• Not for reading a Figma design into code — that's figma-codegen.

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.

First, understand the environment, then build (provider-first)

The write-side mirror of codegen's grounding, and the most important habit: an off-looking build almost always comes from invented values or assumed conventions. So understand this user's actual environment first, then build into it — never apply a generic template. There's no single "right" stack; there's their stack. Look at both ends before creating anything:

• The Figma file — its existing design system, which you'll reuse and bind to:
  1. get_variable_defs → the file's variables (colour / spacing / radius / typography) with names + values + hex. These are the tokens you bind to.
  2. scan_components / get_local_components → existing components to instance rather than rebuild. Match the source UI pattern (a card, a list row, a nav, a button) to a component.
  3. get_styles → shared paint / text / effect styles to apply.
  4. A design system that lives in a published library — these reads list only the file's own definitions. When they come back empty or sparse but the file is built on a team library (its components are library instances, or the user says so), the tokens are in the library file: read their keys there and import what you bind. Don't conclude "no tokens" and hardcode values. → references/write-rules.md, "Tokens from a published library".
• The source you were handed — when it's code, which stack and styling system (Tailwind / Chakra / MUI / CSS modules / vanilla …) and whether it has a config / theme / tokens file. That's where its real values live — read them from there, don't assume a default.

Then build, tracing every value to a source in priority order: reuse an existing component / variable / style; else take the exact value from the source's own code — provider-first, resolving whatever styling system it uses to real px / hex (CSS literally; Tailwind / Chakra / UnoCSS / … via their config or scale), never eyeballed; only invent from a consistent scale when neither exists. Full detail + the value-resolution method in references/write-rules.md.

Two jobs — both follow the write rules

Every build obeys the same cross-cutting rules — ground values (design system → source code → scale), reference tokens (colour via bind_variable_to_paint, scalars via bind_variable_to_node, shared looks via apply_style_to_node), auto-layout for related children (absolute only for top-level placement), HUG/FILL/FIXED sizing via set_layout_props (so the layout computes sizes — don't hardcode width/height), and real fonts (a new TEXT node defaults to Inter). → references/write-rules.md.

• Assemble a screen / component from what exists (the common case): recognise the UI pattern, create_instance matching components, bind tokens, build incrementally, screenshot-verify each step. → references/assemble-screens.md.
• Author a new design-system asset (only when grounding found no equivalent): create variables/collections, paint/text styles, or components + variant sets, then switch back to the reuse path. → references/author-design-system.md.

Motion (animation)

When the source carries animation — CSS @keyframes / transition, Framer Motion, GSAP, a Vue/Svelte transition — author it as Figma Motion (beta) on the frame you built rather than dropping it: apply_animation_style (presets from get_motion_styles) or apply_manual_keyframe_track per property, set_timeline_duration for length. A staggered row is one atomic batch of apply_animation_style ops with increasing config.timelineOffset — not N calls. Motion is Figma-Design-only and keyframes attach to a top-level frame's layers, so build the frame first. → references/motion.md.

Verify visually (close the loop)

get_screenshot the built node, fix discrepancies, and re-screenshot — the same render-and-diff discipline codegen uses, in reverse. Check it against the source intent and against objective design health (this catches problems even when you built from a vague description with no source to compare): nothing clipped or overflowing, edges aligned, spacing consistent (one scale), a clear type hierarchy. An empty: true export means the node rendered nothing (hidden / off-canvas).

Rules

• Understand the environment, then build. Read both the file's design system and the source's stack / styling system before creating anything — there's no single right stack, only theirs. Ground every value (design system → the code's own values → a sensible scale); never invent.
• Reuse beats regenerate. Instance existing components; bind existing variables/styles. Build new only what the system lacks, and name/structure it to fit.
• Reference tokens, not literals. Colour via bind_variable_to_paint, scalars via bind_variable_to_node, shared looks via apply_style_to_node — never hardcode hex/px when a token exists (get_variable_defs tells you what does).
• A colour from code is sRGB — say so. Write it with space: "SRGB" beside its r/g/b, or a Display P3 file paints it more saturated than the code (references/write-rules.md).
• Auto-layout for related children, absolute coordinates only for top-level placement. Centre with the parent's alignment (primaryAxisAlignItems / counterAxisAlignItems; a child that aligns differently gets its own wrapper frame), never by tuning x/y; align_nodes aligns/distributes what has no layout, set_text_properties takes textAlignHorizontal / textAlignVertical. Recipes, shapes (create_shape, boolean_nodes, flatten_nodes) and what the tools still can't draw, in references/write-rules.md "Centering and alignment".
• One create call, fully styled. create_frame, create_rectangle, create_ellipse, create_shape and create_text take fills, strokes, layoutSizing* (frame and rectangle also cornerRadius; a frame also autoLayout and clipsContent; text its typography incl. textAlignHorizontal) — each meaning what it means on its set_* tool. Use them instead of a create plus a chain of follow-ups.
• ok: true is not proof. Confirm layout-sensitive writes by screenshot or read-back, and report anything the tools could not express instead of claiming it done.
• Build incrementally and validate (screenshot) — recognise the UI pattern and assemble it from the matching components, don't reproduce it from primitives.
• Match the file's conventions — naming, structure, and the design system's own patterns, the way codegen mirrors the project's existing code style.