The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external websites, GitHub repositories, and local documents (PDF, PPTX), which constitutes a potential attack surface where embedded instructions in those assets could attempt to steer the agent's behavior.
Ingestion points: WebFetch is used for live website analysis; local file system access is used for reading PPTX, PDF, and source code files.
Boundary markers: The skill does not explicitly define specific delimiters or "ignore instructions" wrappers for the content it ingests during the synthesis phase.
Capability inventory: The skill primarily performs file reads and data synthesis; it delegates complex transformations and uploads to secondary specialized skills.
Sanitization: The instructions recommend user-driven anonymization of sensitive data (like sales figures in PPTX) but do not describe automated sanitization for prompt-based attacks within the assets.
[COMMAND_EXECUTION]: The skill utilizes standard filesystem commands to inspect and analyze the provided design assets.
Evidence: Uses ls for directory structure analysis and Read for accessing file contents in the local environment.

design-system-prep