Labsco
bitwarden logo

using-figma

121

by bitwarden · part of bitwarden/ai-plugins

Read and inspect Figma designs via the Dev Mode MCP server — selects the right tool, parses Figma URLs into fileKey and nodeId, and turns design context into useful input for critique, copy review, handoff prep, and Design System work.

🧩 One of 7 skills in the bitwarden/ai-plugins package — works on its own, and pairs well with its siblings.

This is the playbook your agent receives when the skill activates — you don't need to read it to use the skill, but it's here to audit before installing.

Using Figma via the Dev Mode MCP Server

This skill grounds the designer in the Figma Dev Mode MCP server — Figma's official MCP that exposes design context, variables, screenshots, metadata, and design-system search to Claude. Apply it whenever a task needs to read a Figma design — extract structure, tokens, screenshots, or strings. Composing skills like design-review, content-style-guide, preparing-design-handoff, and evolving-design-system-components call into here whenever a Figma file is referenced.

The Figma MCP server also exposes write tools (creating files, uploading assets, generating diagrams, mapping Code Connect, mutating Figma objects). Those are intentionally out of scope for this skill — its allowed-tools is read-only. If a task genuinely needs a Figma mutation, invoke the underlying MCP tool directly with explicit user consent rather than extending this skill's surface.

Anatomy of a Figma URL

Every interaction starts from a Figma URL. The two pieces that matter are the fileKey and the nodeId.

https://www.figma.com/design/<fileKey>/<fileName>?node-id=<nodeIdWithDashes>&...

Conversion: the URL's node-id uses - as the separator (123-456), but the MCP server expects : (123:456). Convert before passing it into a tool.

If the user pastes a URL without a node-id, the URL points at the whole file. Ask which frame they mean before extracting anything — operating on the whole file is rarely what's wanted and returns far too much context.

The read tools, by job to be done

The Figma MCP server exposes many tools. Pick the smallest one that answers the question.

Job to be doneToolNotes
Read a frame's design context (code + screenshot + metadata)get_design_contextMost common entry point. Accepts a framework parameter; defaults to React + Tailwind.
Get just the structural outlineget_metadataSparse XML of layer IDs, names, types, positions, sizes. Cheap.
Get just the rendered imageget_screenshotVisual reference without the code or token noise.
Extract design tokens used in selectionget_variable_defsVariables and styles — colors, spacing, typography.
Discover available libraries on the fileget_librariesShows which subscribed/available design libraries are linked.
Find a component in the design systemsearch_design_systemText query against components, variables, styles.
Inspect a FigJam boardget_figjamSame role as get_metadata but for FigJam content.
Identify the authenticated Figma userwhoamiUseful when permission / seat type matters.

Code Connect tools (read-only)

These map Figma components to their code counterparts. Mostly relevant inside evolving-design-system-components; rarely needed for critique or copy review.

  • get_code_connect_map — returns existing mappings, source files, and snippets for selected instances.
  • get_code_connect_suggestions — suggested mappings for selected components.
  • get_context_for_code_connect — property definitions and variant options for a component.

(The two write Code Connect tools — add_code_connect_map and send_code_connect_mappings — are out of scope for this skill. Invoke them directly with explicit user consent if needed.)

Decision rules

  • Don't reach for get_design_context by reflex. It returns code and screenshots even when the question is "what color is this background" — get_variable_defs answers that in a fraction of the context.
  • Start with get_metadata for orientation. When the goal is "tell me what's in this frame", the metadata XML is cheaper than full context and usually enough to pick the next move.
  • Use get_screenshot for human reference, get_metadata for machine reasoning. Don't load both unless both are needed.
  • Don't generate code from this skill. Production code generation belongs in repo-specific output skills like figma-to-angular in the clients repo. This skill stops at extracted design context.

Composing with other skills

  • design-review. When critiquing a design, start with get_screenshot + get_metadata to orient. Pull get_variable_defs if tokens are part of the critique (off-system colors, inconsistent spacing). Only escalate to get_design_context when the code-shape itself is the question.
  • content-style-guide. Use get_design_context or get_metadata to surface every user-visible string in a frame, then walk each string through the style guide. Do not rewrite Figma copy from inside this skill — return findings with proposed alternatives and let the designer apply.
  • preparing-design-handoff. Use get_metadata to verify the file has the expected Ready-for-Dev sections with names aligned to engineering stories. Use get_variable_defs to confirm tokens are library-bound rather than raw hex.
  • evolving-design-system-components. Use search_design_system and get_libraries before proposing a new pattern — most "we need this new thing" cases turn out to be "this thing exists in the library and we didn't know." Use the Code Connect read tools when the question crosses into how a Figma component maps to its code counterpart.

Asking the user before extracting

Before fetching, surface what you're about to ask for and why:

"I'll pull get_metadata for that frame first to see the layer structure, then get_variable_defs for the tokens it uses — that should answer the spacing question without loading the full design context. OK?"

This is faster than apologizing for an over-broad call later and helps the designer learn which MCP tool answers their kind of question.

Output format for Figma extractions

When reporting what's in a Figma file, structure the response as:

  1. Frame and stage — file name, frame name, stage if known (30/60/90).
  2. Structure — layer outline at the level relevant to the question (don't paste raw XML).
  3. Tokens and library bindings — what's bound to the design system vs. what's a raw value.
  4. User-visible strings (when copy is part of the task) — flagged with content-style-guide findings.
  5. Open questions — anything ambiguous in the file that the designer should clarify before the work moves forward.

Additional resources

  • references/setup.md — installing and authenticating the Figma Dev Mode MCP server (desktop vs. remote), seat-type requirements, and troubleshooting unavailable tools.
  • Figma's canonical per-tool reference: developers.figma.com/docs/figma-mcp-server/tools-and-prompts/ — the source of truth for tool parameters, return shapes, and desktop-vs-remote availability.