The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes several external command-line utilities including yt-dlp, ffmpeg, pandoc, and various Whisper transcription engines (mlx-whisper, whisper.cpp, openai-whisper). Technical review of the scripts (scripts/fetch_transcripts.py, scripts/list_channel.py, scripts/render_pdf.py) confirms that these are executed using list-based arguments in subprocess.run calls. This implementation avoids the use of shell=True, effectively preventing shell command injection vulnerabilities despite handling user-provided channel URLs and IDs.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from YouTube (video titles, descriptions, and transcripts) which are then used by the agent to synthesize a research document. This presents a potential surface for indirect prompt injection.
Ingestion points: YouTube video metadata and transcripts fetched during Step 2 and Step 4 of the workflow.
Boundary markers: The skill does not implement specific delimiters or 'ignore' instructions for the processed content; the agent is simply instructed to read the transcripts and extract relevant information.
Capability inventory: The skill has the ability to write files to the current working directory, execute media processing tools, and delete the hidden workspace folder (./.youtube-research/) upon user confirmation.
Sanitization: Descriptions are truncated to a specific length in scripts/list_channel.py, and VTT files are parsed into plain text, but no specific sanitization for malicious LLM instructions is performed.

youtube-research