The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external content (YouTube transcripts and thumbnails) which represents an indirect prompt injection surface.
Ingestion points: Transcripts and thumbnail metadata are fetched by scripts/fetch.py.
Boundary markers: The skill uses structured JSON schemas in its prompts and explicitly instructs the model to provide verbatim quotes.
Capability inventory: The skill can execute local Python scripts and write files to the ~/youtube-reports/ directory.
Sanitization: The scripts/cache.py script includes a verify-quotes routine that performs substring-matching to ensure the model's output is grounded in the original transcript data.
[DYNAMIC_EXECUTION]: The script scripts/doctor.py utilizes dynamic module loading.
Evidence: The check() function uses __import__(import_name) to verify the installation of the yt_dlp and youtube_transcript_api libraries. This is a standard and benign implementation for dependency checking.

youtube-verdict