The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to download a desktop application binary from an external GitHub repository (github.com/grab/TalkToFigmaDesktop/releases).
[PRIVILEGE_ESCALATION]: The installation instructions explicitly guide the user to bypass operating system security features. For macOS, it directs users to bypass Gatekeeper using 'Right-click app → Open → Open'. For Windows, it directs users to bypass SmartScreen using 'More info → Run anyway'. These instructions facilitate the execution of untrusted code by circumventing built-in security protections designed to prevent the installation of unauthorized software.
[COMMAND_EXECUTION]: The MCP configuration setup requires the AI client to execute a local file (mcp-server.cjs) via node. This file is dynamically installed or managed by the external desktop application, creating a vector for local code execution controlled by the external software.
[INDIRECT_PROMPT_INJECTION]: The skill retrieves and processes data from external Figma design files, which creates an attack surface for instructions embedded in design layers or properties that could influence the AI's behavior.
Ingestion points: Node data and properties retrieved via tools such as get_node_by_id, search_nodes, and get_all_nodes from Figma files.
Boundary markers: None identified; there are no instructions to the agent to treat content from Figma as untrusted data or to separate it from system instructions.
Capability inventory: The skill has significant capabilities, including creating nodes (create_rectangle, create_text), modifying properties (set_properties), and deleting nodes (delete_node).
Sanitization: No evidence of data validation or sanitization for Figma design content before it is processed by the skill.

talktofigma-desktop-mcp