The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external, untrusted content from YouTube transcripts which could contain hidden instructions for the AI.
Ingestion points: The script scripts/get_transcript.py reads subtitle content fetched from external YouTube URLs using yt-dlp.
Boundary markers: The instructions in SKILL.md do not provide delimiters or warnings to the agent to ignore instructions embedded within the transcript text.
Capability inventory: The skill possesses the capability to execute system commands via subprocess.run (to invoke yt-dlp) and read local files.
Sanitization: The clean_vtt function in scripts/get_transcript.py removes formatting tags and timestamps but does not filter for natural language instructions or malicious prompts.
[COMMAND_EXECUTION]: The skill executes system commands to interact with the underlying OS.
Evidence: The script scripts/get_transcript.py uses subprocess.run to execute the yt-dlp binary. It correctly uses a list-based command structure rather than a shell string, which mitigates standard shell injection risks.
[EXTERNAL_DOWNLOADS]: The skill relies on and fetches data from external sources.
Evidence: The skill requires the installation of the yt-dlp utility via brew or pip as specified in the SKILL.md metadata. It also fetches subtitle files from YouTube's infrastructure during runtime.

youtube-watcher