The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/test_read_marginalia.py uses subprocess.run to execute the skill's own driver script (read_marginalia.py) for automated integration testing. This usage is restricted to testing local CLI functionality using the current Python interpreter and does not pose a risk of arbitrary command injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from multiple sources, including Word document comments (scripts/read_marginalia.py), audio/video transcripts (references/transcript-mode.md), and OCR data from handwritten notes (references/photo-mode.md).
Ingestion points: External .docx files, .vtt transcripts, and image files for handwritten notes.
Boundary markers: The instructions explicitly demand an 'Interview before drafting' phase where the agent must batch clarify questions and verify all extracted data with the user.
Capability inventory: The skill writes Markdown and Word documents to the local file system and appends data to an Excel workbook (scripts/update_tracker.py).
Sanitization: Human-in-the-loop verification is the primary control; the agent is instructed to 'Never guess or correct a name' and to flag all unverified data with warning emojis for user confirmation.

meeting-notes