The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external binaries (ffmpeg, ffprobe, yt-dlp) to process video data.\n
All calls use structured argument lists via subprocess.run, which effectively mitigates shell injection risks.\n- [EXTERNAL_DOWNLOADS]: On macOS systems, the setup script can automatically install required dependencies using Homebrew.\n
It also downloads video content from user-provided URLs using yt-dlp.\n- [DATA_EXFILTRATION]: Extracted audio data is sent to external APIs for transcription.\n
Audio is uploaded to api.groq.com or api.openai.com depending on configuration.\n- [CREDENTIALS_UNSAFE]: The skill manages API keys for transcription services using a local .env file (~/.config/watch/.env).\n
It implements proper security measures by setting restricted file permissions (chmod 600) and avoiding hardcoded secrets.\n- [INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data from video sources and transcripts, representing an indirect prompt injection surface.\n
Ingestion points: Video captions fetched via yt-dlp or transcripts generated via Whisper API in scripts/watch.py.\n
Boundary markers: Transcripts are enclosed in markdown code blocks in the final report header.\n
Capability inventory: Execution of ffmpeg, ffprobe, and yt-dlp via subprocess.run across multiple scripts.\n
Sanitization: Arguments are passed to subprocesses as lists; yt-dlp specifically uses the -- separator to prevent flag injection from untrusted URLs.

watch