The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructions specify a 'Step 3-delta' escape hatch where the agent generates and executes Javascript code using the figma_execute or use_figma MCP tools. This logic is triggered when data needed for the accessibility spec is missing from the local cache. While described as read-only, this constitutes runtime code generation based on potentially untrusted input.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from optionalContext (free-form user text) and _base.json (external plugin output). This data is used to infer behavioral states and guide the reasoning process.
Ingestion points: The optionalContext variable and the {cachePath}/{componentSlug}-_base.json file.
Boundary markers: No specific delimiters or safety instructions are defined to separate untrusted context from internal logic.
Capability inventory: The skill can write JSON files to the local file system and execute dynamic Javascript within the Figma environment via MCP tools.
Sanitization: No explicit sanitization or validation of the optionalContext or JSON data is described before it is used in logic or output generation.
[COMMAND_EXECUTION]: The skill utilizes MCP tools (figma_execute and use_figma) to run environment-specific commands (Javascript in Figma) based on its own generated logic, which can be influenced by external input.

extract-voice