The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to parse and process untrusted document formats (PDF, DOCX, XLSX, PPTX), which serve as a vector for indirect prompt injection and other document-based attacks.
Ingestion points: Documents uploaded or provided by users are processed in SKILL.md and references/LIBRARIES.md for text extraction and manipulation.
Boundary markers: While no explicit prompt delimiters are provided for raw text extraction, the skill mandates autoescape=True when rendering document data via Jinja2 templates to prevent injection into the output document.
Capability inventory: The skill utilizes Write, Edit, and Bash tools (specifically for libreoffice and qpdf) to handle file system operations and document conversions.
Sanitization: The skill provides a safe_load implementation to enforce size and page limits and explicitly warns developers to treat documents as hostile input, highlighting risks like macros and XXE.
[COMMAND_EXECUTION]: The skill provides a implementation recipe that uses subprocess.run to execute external binaries for document conversion.
Evidence: The convert function in references/LIBRARIES.md invokes libreoffice via subprocess.run to perform headless document format conversion.

document-processing-expert