The Agent Skills Directory

[DATA_EXPOSURE_AND_EXFILTRATION]: The skill requires access to an API key stored at ~/.config/typesafe/typesafe-key.header. The instructions explicitly direct the agent to check for the file's existence without reading its contents into the conversation, preventing accidental leakage. The key is only used locally as a header for curl requests to the official vendor API (api.typesafe.ai).
[COMMAND_EXECUTION]: The skill executes local shell commands including node to run its internal pack-requests.mjs script and curl to interact with the TypeSafe API. These are necessary for the skill's stated purpose of processing data and communicating with the model.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it processes external content from Figma designs (spec notes, node IDs) and source code.
Ingestion points: Figma node data, spec notes, and project source files (SKILL.md, Step 2).
Boundary markers: Uses structured JSON formats and markdown headers to separate data, and requires English 'glosses' for translated text.
Capability inventory: File system access for reading sources and writing to a temporary run folder, and network access via curl to the vendor API.
Sanitization: The pack-requests.mjs script performs strict JSON validation and image header parsing but does not sanitize the text content of questions or facts against adversarial instruction patterns.

ask-jev