The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from potentially untrusted Word documents. Malicious documents could contain instructions designed to influence the agent's behavior during content analysis.
Ingestion points: Content is imported into the context via ooxml/scripts/unpack.py and pandoc extraction commands described in SKILL.md.
Boundary markers: There are no explicit instructions for the agent to use delimiters or ignore embedded instructions within processed document text.
Capability inventory: The skill can execute shell commands (soffice, git, pandoc) and perform file system writes via scripts/document.py and ooxml/scripts/pack.py.
Sanitization: The implementation consistently uses the defusedxml library, which effectively mitigates XML External Entity (XXE) and other XML-based injection vulnerabilities.
[COMMAND_EXECUTION]: The skill executes system commands to perform document validation and comparison tasks. ooxml/scripts/pack.py uses subprocess.run to call soffice for converting and validating Office files, and ooxml/scripts/validation/redlining.py uses it to call git diff for comparing document revisions. These operations are limited to the skill's primary document manipulation functions.

docx