The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied documents in Word (.docx), Excel (.xlsx), PowerPoint (.pptx), and PDF formats. These files are processed through various libraries and tools, creating a vector for instructions embedded within documents to influence the agent.
Ingestion points: The skill ingests external data through libraries like pandas and openpyxl (in xlsx/recalc.py), pypdf and pdfplumber (in pdf/SKILL.md), and CLI tools such as pandoc (in docx/SKILL.md) and markitdown (in pptx/SKILL.md).
Capability inventory: The skill environment includes the ability to execute shell commands (via subprocess.run calls to soffice, pandoc, qpdf, and pdftoppm) and perform file system write operations.
Boundary markers: The instructions do not specify explicit boundary markers or delimiters to isolate data from instructions when reading external documents.
Sanitization: The skill documents use of defusedxml for secure XML handling, which helps mitigate standard XML-based attack vectors.
[DYNAMIC_EXECUTION]: The xlsx/recalc.py script generates a LibreOffice Basic macro (Module1.xba) and writes it to the user's local configuration directory (~/.config/libreoffice/ or ~/Library/Application Support/LibreOffice/) at runtime. This script generation and execution pattern is a form of dynamic code execution.
[COMMAND_EXECUTION]: The skill instructions and scripts frequently invoke external system utilities for document conversion and analysis. These include LibreOffice (via the soffice command), pandoc, qpdf, and the poppler-utils suite (pdftotext, pdftoppm).

document-skills