The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on executing system binaries for its core functionality. It uses the Python subprocess module to call ffmpeg and ffprobe for media processing, yt-dlp for downloading content, and the brew package manager on macOS for dependency installation. These commands are constructed as lists to prevent shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of video content and subtitles from external URLs provided by the user via yt-dlp. It also automates the installation of required binaries (ffmpeg, yt-dlp) through system package managers (brew, apt, winget). Additionally, it utilizes the --remote-components feature of yt-dlp to fetch official JavaScript solvers from GitHub to handle YouTube's platform challenges.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of video transcripts and captions.
Ingestion points: Transcripts are fetched from video platforms (via scripts/download.py) or generated via the Deepgram API (via scripts/deepgram.py) and processed in scripts/watch.py.
Boundary markers: The skill wraps the resulting transcript in markdown code blocks when outputting the report to the agent.
Capability inventory: The agent has access to Bash (used to run the skill's scripts) and the Read tool (used to view extracted JPEG frames).
Sanitization: The scripts/transcribe.py script performs basic sanitization by stripping HTML and WebVTT tags from the content using regular expressions.
[SAFE]: The skill implements security best practices for secret management. It stores the Deepgram API key in a local .env file within a dedicated configuration directory (~/.config/watch/) and explicitly sets restricted file permissions (chmod 600) to ensure the key is not readable by other users on the system. All network communications for transcription are directed to Deepgram, a well-known service provider.

watch