The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes system tools such as soffice (LibreOffice), pandoc, and pdftoppm via subprocess calls to perform document conversions and rendering. These calls use argument lists which mitigate shell-based command injection risks.\n- [COMMAND_EXECUTION]: In scripts/accept_changes.py, the skill dynamically generates a LibreOffice Basic macro string and writes it to a configuration file in a temporary profile directory before executing it. This technique is used to programmatically automate the 'Accept All Tracked Changes' functionality in LibreOffice headless mode.\n- [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it is designed to ingest and process data from external Word (.docx) and PDF files.\n
Ingestion points: External documents are parsed via pdf2docx and scripts/unpack.py (which unzips and reads the underlying XML parts).\n
Boundary markers: The skill instructions do not specify the use of boundary markers or 'ignore' instructions to isolate document content from the agent's operational logic.\n
Capability inventory: The skill has access to subprocess execution and file system writes, which could be triggered by instructions embedded in a processed document.\n
Sanitization: There is no evidence of content sanitization or filtering to remove potential instructions from the extracted document data.

docx