The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/understand_video.py executes system binaries ffmpeg and ffprobe using the subprocess.run method. The script correctly passes arguments as a list rather than a shell string, effectively mitigating shell injection risks.
[EXTERNAL_DOWNLOADS]: The script includes logic to detect if the openai-whisper dependency is missing and offers to install it from PyPI using pip. PyPI is a well-known and trusted package registry.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from video files which may contain embedded instructions designed to influence the agent.
Ingestion points: The scripts/understand_video.py script ingests metadata and audio content from user-provided video files, which are then parsed by ffprobe and Whisper respectively.
Boundary markers: The script outputs results in a structured JSON format. However, the text and transcript fields contain raw data from the video without specific delimiters or warnings to the agent to ignore instructions embedded within the content.
Capability inventory: The skill possesses the capability to read and write files in the directory where the video is located and to execute the ffmpeg and ffprobe binaries.
Sanitization: No sanitization or filtering is applied to the transcribed text or video metadata before it is returned to the agent.

video-understand