The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted Office documents and HTML slide content, which presents a surface for indirect prompt injection attacks. \n
Ingestion points: ooxml/scripts/unpack.py (extracts XML from office files), scripts/inventory.py (extracts text content), and the html2pptx workflow (reads user-provided HTML). \n
Boundary markers: Absent. Extracted content is provided directly to the agent without specific delimiters or instructions to treat it as untrusted. \n
Capability inventory: The skill can execute shell commands via subprocess.run (soffice, pdftoppm, git), write files to the workspace, and potentially perform network operations. \n
Sanitization: Employs defusedxml in ooxml/scripts/unpack.py and ooxml/scripts/pack.py to mitigate XXE risks.\n- [COMMAND_EXECUTION]: Core functionality relies on executing system binaries via Python's subprocess module. \n
Evidence: ooxml/scripts/pack.py and scripts/thumbnail.py invoke soffice for converting presentations to PDF and HTML formats. scripts/thumbnail.py invokes pdftoppm for slide image generation. ooxml/scripts/validation/redlining.py uses git diff for comparing document versions. \n
Note: These executions use fixed binary names and path-based arguments for standard document processing tasks.\n- [DYNAMIC_EXECUTION]: The presentation creation workflow involves the agent generating and running temporary driver scripts. \n
Evidence: SKILL.md (Step 6 of the html2pptx workflow) instructs the agent to write a JavaScript file using the @ant/html2pptx library and execute it with node. \n
Note: While the script logic is AI-generated, it processes external data (HTML slides) which could influence the behavior if not handled carefully.\n- [EXTERNAL_DOWNLOADS]: The skill requires several external packages and system tools. \n
Dependencies include markitdown, pptxgenjs, playwright, and a local package html2pptx.tgz. Installations are from standard registries or local sources bundled with the skill.

pptx