The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from YouTube, including video titles and transcripts, which are stored in local Markdown files and subsequently read by the agent.
Ingestion points: scripts/download.sh fetches metadata and subtitles using yt-dlp; scripts/vtt_to_transcript.py processes the resulting VTT files into plain text.
Boundary markers: Transcripts are placed under a ## Transcript header in the generated Markdown files. While this provides structure, there are no explicit 'ignore embedded instructions' warnings for the AI agent.
Capability inventory: The skill uses subprocess.run to execute ffmpeg and yt-dlp in scripts/extract_slides.py and scripts/contact_sheet.py. It also includes a local web server (scripts/serve.py) that can modify Markdown files via PATCH requests.
Sanitization: scripts/vtt_to_transcript.py includes a markdown_text function that escapes special markdown characters and HTML entities to ensure the ingested text remains inert.
[COMMAND_EXECUTION]: The skill relies on external command-line utilities to perform video and image processing tasks.
Evidence: scripts/extract_slides.py and scripts/contact_sheet.py use subprocess.run to invoke ffmpeg. The arguments for these calls, such as timestamps, are validated as floats or rounded numbers to prevent injection.
Evidence: Shell scripts like scripts/download.sh and scripts/detect_slides.sh invoke yt-dlp and ffmpeg. Input YouTube IDs are validated against a strict 11-character regular expression.
[EXTERNAL_DOWNLOADS]: The skill downloads media and metadata from external servers as part of its core functionality.
Evidence: scripts/download.sh utilizes yt-dlp to fetch video files and subtitles from YouTube.
Evidence: scripts/setup.sh makes network requests to the YouTube oembed API to verify if a video allows embedding.

youtube-notetaker