The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses yt-dlp to download video content from user-provided URLs as a mandatory first step for remote sources.
Evidence: yt-dlp -f "bestvideo[height<=720]+bestaudio/best[height<=720]" --merge-output-format mp4 -o "$OUT/source.mp4" "$URL" in SKILL.md.
[COMMAND_EXECUTION]: The skill executes multiple shell commands to process video data, including ffprobe for metadata extraction, whisper for audio transcription, and ffmpeg for keyframe generation.
Evidence: ffprobe -v error -show_entries ... "$VIDEO", whisper "$VIDEO" ..., and ffmpeg -ss ... -i "$VIDEO" -vframes 1 ... in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted video and audio data which is then processed by an LLM (Gemini) with instructions to perform content-level analysis.
Ingestion points: Video files retrieved from local paths or external URLs via yt-dlp (SKILL.md).
Boundary markers: The instructions do not define specific delimiters or "ignore embedded instructions" warnings for the multimodal analysis step (SKILL.md).
Capability inventory: The skill utilizes shell execution for downloading, probing, transcribing, and editing video files, as well as file system access for generating reports (SKILL.md).
Sanitization: No explicit validation or sanitization of user-provided URLs or video-derived metadata used in shell command interpolation is documented (SKILL.md).

video-analysis