The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content from video files and YouTube, creating a potential attack surface for indirect instructions.\n
Ingestion points: Tools such as video_info and video_watch process data from external YouTube URLs and local files, specifically targeting transcripts and captions (SKILL.md).\n
Boundary markers: There are no defined delimiters or instructions to treat extracted transcription text as untrusted content.\n
Capability inventory: The skill utilizes tools that perform network operations (downloading via yt-dlp) and file system access to extract frames and metadata. It also includes a video_setup tool for installing dependencies.\n
Sanitization: The workflow does not include steps to sanitize or validate the content of video transcripts before the agent incorporates them into its reasoning context.

video-perception