The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/docx_quality_gate.py uses subprocess.run to invoke libreoffice or soffice for a conversion smoke test. The command is constructed using a structured list of arguments rather than a shell string, which mitigates the risk of command injection.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process and extract data from Word documents, which is a potential surface for indirect prompt injection.
Ingestion points: Untrusted document content is ingested via scripts/docx_extract.py and scripts/docx_to_html.mjs.
Boundary markers: The documentation in SKILL.md and references/llm-extraction-workflows.md warns the agent to treat macro-enabled files as untrusted and provides explicit guidance on ignoring instructions inside documents.
Capability inventory: The skill has Bash and Write capabilities, which could be misused if the agent obeys instructions found within a processed document.
Sanitization: The skill identifies that mammoth.js does not sanitize HTML output and explicitly instructs the user to sanitize it before use, even embedding a security warning into the generated HTML files.
[EXTERNAL_DOWNLOADS]: The skill references various libraries and documentation sources in data/sources.json, including those from Microsoft, IBM (Docling), and Mistral AI. These are established technology vendors and their official resources are considered trusted.

document-docx