The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external .docx files provided by users. While this presents an attack surface for indirect injection via document content, the skill effectively mitigates common technical exploits.
Ingestion points: XML content from unpacked .docx archives is processed in document.py and utilities.py.
Boundary markers: The skill relies on programmatic XML parsing rather than raw string interpolation into LLM prompts.
Capability inventory: File system operations (unzip/pack), subprocess execution for document conversion (soffice), and diffing (git).
Sanitization: The skill consistently uses defusedxml for XML parsing in document.py, utilities.py, pack.py, and unpack.py, which is the industry standard for preventing XML External Entity (XXE) and billion laughs attacks.
[COMMAND_EXECUTION]: The skill utilizes subprocess.run in ooxml/scripts/pack.py to call soffice for PDF conversion and in ooxml/scripts/validation/redlining.py to call git diff for change validation. These executions are limited to specific, functional paths and do not accept arbitrary user-controlled shell input.
[EXTERNAL_DOWNLOADS]: Dependencies listed in SKILL.md (pandoc, libreoffice, poppler-utils, defusedxml, docx) are well-known, established tools from trusted repositories and registries.

docx