The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted video files and metadata from social media platforms.
Ingestion points: TikTok/Reels URLs and the corresponding downloaded video files are processed in scripts/run_video_analysis_batch.mjs.
Boundary markers: The Gemini prompt construction in buildPrompt does not utilize delimiters or specific instructions to the model to ignore potential instructions embedded in the video's content.
Capability inventory: The skill has permissions to execute shell commands (ffprobe, yt-dlp), read/write local files, and perform network requests to a hosted API.
Sanitization: No sanitization of video data or source metadata is performed before submission to the language model.
[COMMAND_EXECUTION]: The skill executes external tools yt-dlp and ffprobe via Node.js spawn and spawnSync. While these are necessary for the skill's stated purpose, they represent the execution of external binary code.
[EXTERNAL_DOWNLOADS]: The skill's core workflow involves downloading video content from external TikTok and Reels URLs.

video-analysis