figma

Figma evidence profile

Type: System profile
Status: Active
Canonical for: Factual interpretation of preserved Figma MCP and Code Connect evidence
Depends on: corpus/figma/

Role in the field

Figma is not a design system in this comparison. It is a bridge that exposes design-file structure and design-to-code mappings to agents. It therefore tests how much “design context” can become machine-readable and where the agent still supplies judgment.

Four-question relevance

• Intent: conditional on custom rules/instructions; geometry and mappings do not explain why a component was chosen.
• Structure: demonstrated for design/code identity, properties, variants, examples and some nesting.
• Retrieval: demonstrated through MCP context and Code Connect snippets.
• Constraint/evaluation: demonstrated for mapping integrity, not semantic fitness.

Machine-readable design context

• Documented: Figma MCP can return design context, screenshots, metadata, variable definitions, Code Connect mappings, library search results and other file information through named tools. Tools and prompts
• Documented: Figma distinguishes what MCP sends from what the agent does. The server provides structured context; the agent interprets and generates. MCP versus agent
• Documented: Files should be structured for retrieval, large frames avoided, variables mapped to code, and custom rules/skills supplied for codebase-specific behavior. Figma MCP documentation set
• Documented: create_design_system_rules helps create a portable rules file that tells agents how to translate designs into codebase-aware frontend output. Tools and prompts

Code Connect

• Documented: When Code Connect is present, MCP output includes actual implementation details and custom instructions rather than only visual structure. Code Connect integration
• Implemented: The public Code Connect repository supports component connections, prop mappings, variant restrictions, examples, imports, nested props/slots and multiple frameworks. Code Connect snapshot
• Implemented: Test fixtures show mappings such as figma.enum from Figma variants to code props, boolean mappings, child instances, different examples and variant-restricted connections. Prop mapping fixture, variant restriction
• Implemented: The changelog documents CI behavior, custom parsers, Storybook migration, slot support and parserless templates. Code Connect changelog

What agents can and cannot infer

Strongly machine-readable

• Node hierarchy and selected metadata.
• Variables and code syntax when configured.
• Component-to-code identity.
• Prop/variant mappings and example code.
• Screenshots and visual reference.
• User-authored custom rules.

Still implicit

• Why a component was chosen.
• Whether the composition is appropriate for the task.
• Product-specific hierarchy, density and content strategy.
• Accessibility rationale beyond what is embedded in the mapped implementation/rules.
• Whether a visually similar layer should be replaced by an actual component instance.
• Valid cross-component relationships unless separately documented.

Figma’s own preserved resource guidance warns that MCP cannot understand why a decision was made, replace judgment, or guarantee accurate code from a messy system. MCP for designers

Composition and generated UI

• Documented: Figma supports code-to-canvas, write-to-canvas and design-system search workflows. Code-to-canvas, write-to-canvas
• Inferred: Bidirectional workflow preserves more identity than screenshot-to-code generation, especially with Code Connect. It still does not supply a general composition grammar.

Governance and quality

• Implemented: Code Connect has parsers, validation, tests and CI-oriented failure behavior.
• Documented: Access/rate limits and client/tool failure pages make operational boundaries explicit.
• Gap: Figma MCP exposure can improve grounding but does not itself test semantic fitness, accessibility or task success. Those checks must come from implementations, tests, rules or human review.

Evidence assessment

Figma demonstrates that design metadata and real code mappings materially improve agent grounding. It also provides some of the strongest evidence for the boundary between retrieval and judgment: Code Connect can tell an agent which code corresponds to this design component; it does not tell the agent whether this is the right experience to build.