The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill extracts content from external PowerPoint files, creating a surface for indirect prompt injection.
Ingestion points: Content enters the agent context via text extraction in SKILL.md (using markitdown) and raw XML access via ooxml/scripts/unpack.py.
Boundary markers: Absent. No specific delimiters or instructions to ignore embedded content are defined when extracting text from presentations.
Capability inventory: The skill can execute system commands via subprocess.run (soffice, pdftoppm, git) and render HTML using playwright in scripts/html2pptx.js.
Sanitization: Absent. External text content is processed without explicit sanitization or filtering.
[COMMAND_EXECUTION]: Multiple scripts execute shell commands to support document conversion and validation tasks.
Evidence: scripts/thumbnail.py calls soffice and pdftoppm; ooxml/scripts/pack.py calls soffice; ooxml/scripts/validation/redlining.py calls git.
Context: The shell calls are implemented using argument lists rather than raw string concatenation, which follows security best practices for preventing command injection.

pptx