The Agent Skills Directory

[COMMAND_EXECUTION]: The watch.py script invokes external binaries (yt-dlp, ffmpeg, ffprobe) to process video files and metadata. It utilizes safe execution patterns by passing arguments as a list rather than a single string and using the -- delimiter to prevent user-supplied URLs from being interpreted as command flags.
[DATA_EXFILTRATION]: The skill is designed to upload audio data to external transcription services (Groq and OpenAI) when captions are missing. This is a primary functional requirement documented in the skill instructions. It relies on user-provided API keys stored securely in a local configuration file (~/.config/watch/.env) rather than hardcoded credentials.
[INDIRECT_PROMPT_INJECTION]: As the skill ingests transcripts from external, untrusted video sources, it is subject to indirect prompt injection if the subtitles contain malicious instructions. The skill mitigates this by instructing the agent to treat the transcript as data and by using markdown code blocks to provide structural boundaries.
Ingestion points: Transcripts fetched from YouTube or processed via Whisper are returned as text to the agent's context.
Boundary markers: The script wraps the transcript output in triple-backtick markdown blocks to differentiate it from instructions.
Capability inventory: The skill can execute the watch.py script, perform file operations in a temporary directory, and read frame images.
Sanitization: The script cleans VTT formatting tags from the transcript text but does not filter the natural language content, which is expected for its purpose.

watch