The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external design files which could contain malicious prompts meant to influence the agent.
Ingestion points: Import modules for SVG, Figma (.fig), and REST API JSON described in SKILL.md and evidence.md.
Boundary markers: None. There are no instructions for the agent to disregard text found within design assets.
Capability inventory: High. Includes the ability to execute headless scripts and modify document state via figcraft-cli.
Sanitization: Structural validation and repair are mentioned, but text content is not filtered for prompts.
[COMMAND_EXECUTION]: The skill requires the agent to execute shell commands for design tasks.
Evidence: Use of figcraft-cli subcommands like run, export, and convert is documented throughout SKILL.md.
[EXTERNAL_DOWNLOADS]: Live collaboration and plugin publishing require connecting to remote servers.
Evidence: Mentions of figcraft-server and wss:// protocols in references/automation-surfaces.md.
[DYNAMIC_EXECUTION]: The tool supports a plugin architecture that runs WebAssembly code at runtime.
Evidence: SKILL.md describes a sandboxed WebAssembly environment using the wasmi interpreter for plugins.

figcraft