The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/watch.py uses subprocess.run to execute shell commands (yt-dlp, ffmpeg, ffprobe). While the arguments are passed as a list (which is generally safer), the url parameter is directly incorporated into the command line for yt-dlp.
Evidence: _run(["yt-dlp", "-f", "mp4/best", ..., url]) in scripts/watch.py.
[EXTERNAL_DOWNLOADS]: The skill is designed to download video files and metadata from a wide range of external domains (YouTube, Instagram, X, Vimeo, etc.) using yt-dlp. This involves significant network activity and interaction with untrusted remote servers.
Evidence: Instructions in SKILL.md and README.md to pass arbitrary video URLs to the watch.py script.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (video transcripts and subtitles) and presents them to the AI agent for analysis. This creates a surface for indirect prompt injection where instructions could be hidden in the subtitles of a processed video to influence the agent's behavior during the 'breakdown' phase.
Ingestion points: transcript.txt and index.json generated from external sources in scripts/watch.py.
Capability inventory: The skill utilizes file read/write operations and subprocess execution.
Sanitization: The script performs basic SRT cleaning (removing HTML-like tags) but does not sanitize the text for potentially malicious natural language instructions.
Boundary markers: None identified in the prompt templates provided to the agent in SKILL.md.

watch-video