The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of shell commands (yt-dlp, ffmpeg, ffprobe, awk, python3) to process video files and metadata. These commands incorporate user-provided URLs and file paths, which requires the underlying agent to ensure proper shell escaping to prevent command injection.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, including YouTube descriptions, platform-provided transcripts, and visual descriptions of video frames. This content is subsequently interpolated into prompts for Claude or Gemini, creating a surface for indirect prompt injection if the video content contains malicious instructions designed to influence the agent's behavior.
Ingestion points: Video metadata (descriptions), external transcripts (YouTube/Loom/Riverside), and visual analysis of frames in SKILL.md.
Boundary markers: The skill does not explicitly define delimiters or "ignore instructions" warnings when passing transcript data to the LLM.
Capability inventory: The skill has file system access (~/Documents/videos/), network access (curl to Gemini API, yt-dlp), and shell command execution capabilities.
Sanitization: No explicit sanitization or filtering of external content is mentioned before interpolation.
[EXTERNAL_DOWNLOADS]: The skill uses yt-dlp to download video and audio content from various third-party platforms including YouTube, Loom, Vimeo, X, TikTok, and Instagram. While this is the intended purpose, it involves fetching binary and text data from external, user-specified URLs.
[SAFE]: In multimodal mode, the skill uploads video files to Google's Gemini API for processing. This is a documented feature that targets a well-known service provider and is handled according to platform security guidelines.

watch-video