The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates its workflow by executing local Python scripts (fetch.py, doctor.py, cache.py, segments.py) using subprocess calls. These scripts are provided within the skill's own directory and are used for fetching video metadata, checking dependencies, and managing a local cache.
[EXTERNAL_DOWNLOADS]: The fetch.py script downloads video thumbnails from YouTube's official CDN to the local cache directory (~/youtube-reports/.cache/). These downloads are performed using standard library modules and target trusted domains.
[DYNAMIC_EXECUTION]: The doctor.py script uses the __import__ function to programmatically check for the presence of required Python packages (yt-dlp, youtube-transcript-api). The package names are hardcoded in the script, ensuring only the intended libraries are checked.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data from YouTube.
Ingestion points: Transcripts are fetched from YouTube and passed to the LLM across multiple processing passes.
Boundary markers: The prompts (extract_structure.md, extract_artifacts.md) define rigid JSON output structures and require verbatim quotes from the transcript, which provides a natural boundary against malicious instructions in the content.
Capability inventory: The skill can execute local scripts via subprocess and write report files to ~/youtube-reports/.
Sanitization: The skill strips WebVTT tags and utilizes structured data extraction to minimize the risk of the model interpreting transcript content as instructions.

youtube-extract