The Agent Skills Directory

[COMMAND_EXECUTION]: The main script scripts/understand_video.py performs extensive command execution using ffmpeg and ffprobe to extract metadata and video frames. The implementation uses subprocess.run with argument lists rather than shell strings, which follows best practices for preventing shell injection.
[DYNAMIC_EXECUTION]: In scripts/understand_video.py, the function _check_and_offer_install offers to install the openai-whisper package via pip if it is missing. This operation targets a well-known package from a reputable service and includes an interactive user confirmation prompt, though it technically involves runtime modification of the environment.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of untrusted external data (video files) and converts it into natural language text (transcription) for the agent's consumption.
Ingestion points: Audio content is ingested and transcribed into text in scripts/understand_video.py through the transcribe_with_whisper function.
Boundary markers: The output is structured in a JSON schema which provides clear data boundaries; however, there are no instructions to the agent to treat the transcript or text fields as untrusted content.
Capability inventory: The script has the capability to run local shell commands (ffmpeg, ffprobe) and perform file system writes (creating frame directories and outputting JSON).
Sanitization: No sanitization or filtering is performed on the transcribed text content before it is passed to the agent, potentially allowing an attacker to embed instructions in a video's audio track.

video-understand