The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute system commands for video processing, including ffmpeg, ffprobe, yt-dlp, and whisper. Evidence includes multiple instances in scripts/process_video.py where commands are executed using the recommended list format to prevent shell injection. A utility function in the setup script (scripts/setup.py) uses shell=True but is not invoked in the primary operational logic.\n- [EXTERNAL_DOWNLOADS]: The skill downloads video content from user-specified URLs using the yt-dlp tool. This is a core functionality performed in scripts/process_video.py via the download_video function. The skill also recommends installing standard, well-known Python packages and system utilities from official repositories.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external video data that could potentially contain adversarial instructions designed to influence the AI's behavior.\n
Ingestion points: User-provided video sources (YouTube URLs or local files) processed in scripts/process_video.py.\n
Boundary markers: Absent; the skill does not implement explicit delimiters or warnings for the AI when processing interpolated video data.\n
Capability inventory: The skill utilizes subprocess.run for local tool execution (in scripts/setup.py, scripts/check_providers.py, and scripts/process_video.py) and performs network operations (in scripts/process_video.py) using google.generativeai, openai, groq, assemblyai, and deepgram libraries.\n
Sanitization: Absent; no validation or filtering of content extracted from video files is performed before processing.\n
Note: This represents an inherent risk in multimodal data processing that is mitigated by the safety guardrails of the AI providers used.

video-understand