The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the agent to execute shell commands to render and preview figures. Specifically, it uses npm run svg to process specifications and python3 -m http.server to visualize the output. The rendering command correctly utilizes a quoted heredoc (<<'SPEC') to pass JSON data, which helps prevent shell expansion of the data content.
[INDIRECT_PROMPT_INJECTION]: The skill documentation describes a system where the generated SVGs embed their own specifications in metadata. It notes that a separate tool, the 'docs-skill generator', extracts this metadata to write narration for agents. This creates a multi-step capability chain where untrusted data used to generate a diagram (e.g., descriptions of request flows or background jobs) could result in malicious instructions being injected into the context of agents that later process the generated documentation.
Ingestion points: JSON specifications generated based on user-provided descriptions of system behaviors and request flows.
Boundary markers: The skill uses a quoted heredoc for the initial rendering command, but no boundary markers or sanitization steps are documented for the downstream phase where metadata is converted back into agent instructions.
Capability inventory: The skill utilizes shell command execution (npm, python3) and file system writes (SVG output).
Sanitization: No explicit sanitization or validation of the input data or extracted metadata is mentioned in the instructions.

figure