The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external transcript data and instructs the agent to read and summarize it, creating a vulnerability surface where malicious content in a transcript could influence agent actions.
Ingestion points: The scripts/parse-vtt.py script reads raw VTT files from the filesystem, and SKILL.md (Step 3) instructs the agent to read the generated Markdown files to perform summarization.
Boundary markers: No explicit delimiters or specific instructions are provided to the agent to treat the transcript content as untrusted data or to ignore embedded instructions during the summarization step.
Capability inventory: The skill allows the agent to execute shell commands (via uv run in SKILL.md Step 2) and perform filesystem operations, including moving files (Step 4) and deleting files (Step 5).
Sanitization: The scripts/parse-vtt.py script performs minimal regex-based cleanup of VTT tags but does not sanitize the text content for safe interpolation into the LLM context.
[DATA_EXPOSURE]: The skill instructions direct the agent to access the user's home directory to search for files, which is a sensitive location.
Evidence: SKILL.md Step 1: "Check both meetings/incoming/ and ~/Downloads/ for .vtt files."

process-transcript