The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document content through its extraction tools, which could contain instructions intended to influence the AI agent's behavior.
Ingestion points: The ooxml/scripts/unpack.py script extracts XML content from uploaded .docx files into the workspace.
Boundary markers: Workflow instructions in SKILL.md do not explicitly define boundary markers to isolate extracted document text from the agent's primary instructions.
Capability inventory: The skill ships with scripts that can execute local shell commands (ooxml/scripts/pack.py) and perform file system operations.
Sanitization: The skill mitigates XML-based attacks by using the defusedxml library for all XML parsing operations.
[COMMAND_EXECUTION]: Local shell commands are executed using established system utilities to support the skill's primary document processing operations.
Evidence: ooxml/scripts/pack.py uses subprocess.run to invoke soffice (LibreOffice) for document validation.
Evidence: ooxml/scripts/validation/redlining.py uses subprocess.run to call git diff for comparing document content during the redlining workflow.
Context: These executions are limited to well-known tools and are associates with the intended primary purpose of the skill.

docx