The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands using external tools yt-dlp and ffmpeg to download video data, extract subtitles, and generate screenshots. These commands use external URLs provided by the user.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from YouTube (subtitles, descriptions, and titles) which could contain malicious instructions.
Ingestion points: Video metadata (title, channel, description) and .vtt subtitle files are read by the agent and processed via the local scripts/vtt.py script.
Boundary markers: The instructions direct the agent to "Summarise in your own words" but do not define explicit boundary markers or delimiters to isolate the external video text from the agent's internal instructions.
Capability inventory: The agent has permissions to execute shell commands (yt-dlp, ffmpeg, rm), write files to the vault (Obsidian attachments), and read/display images.
Sanitization: The provided vtt.py script performs text sanitization by stripping HTML tags (re.sub(r"<[^>]*>", "", body)) and decoding entities, but these measures are focused on formatting rather than preventing instruction-based injection attacks.

youtube-transcribe