The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external files such as JSON sales data and PPTX templates and writes the results to the filesystem, creating an attack surface where maliciously crafted data could influence agent behavior.\n
Ingestion points: SKILL.md contains examples that read data from sales_data.json, template.pptx, and presentation.pptx.\n
Boundary markers: The skill does not implement delimiters or provide instructions to the agent to ignore potential commands embedded in the data being processed.\n
Capability inventory: The skill has file-writing capabilities through the python-pptx library (prs.save()) and standard Python file operations (open().write()).\n
Sanitization: No content validation or sanitization is performed on the data ingested from external files before it is processed or used to generate new documents.

powerpoint