The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted YouTube transcripts and metadata which could contain malicious instructions. Ingestion points: YouTube transcripts are fetched via scripts/fetch_transcript.py or supplied by the user as text. Boundary markers: SKILL.md contains explicit instructions for the agent to treat captions, titles, and descriptions as untrusted source material and not as tool instructions. Capability inventory: scripts/fetch_transcript.py uses subprocess.run to call the yt-dlp CLI tool. Sanitization: Instructions in SKILL.md direct the agent to summarize content without executing embedded requests (e.g., Eval 10 tests this behavior).\n- [COMMAND_EXECUTION]: The helper script scripts/fetch_transcript.py uses subprocess.run to execute the yt-dlp CLI tool for metadata extraction. Input video IDs are validated against a strict regex (^[a-zA-Z0-9_-]{11}$) before inclusion in the command to prevent command injection.\n- [DYNAMIC_EXECUTION]: The testing script scripts/test_skill.py dynamically loads the transcript fetching module using importlib for verification purposes. It includes path-validation logic to ensure the module is loaded only from within the own package's directory.

youtube-transcript-dossier