The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted video files provided by users or external sources. A malicious video could contain instructions within its audio or visual frames that, when transcribed or described by the agent, attempt to override the agent's instructions.
Ingestion points: Input video file path and content processed in scripts/understand_video.py.
Boundary markers: Output is formatted as structured JSON, but the transcript text and frame images are raw data that the agent interprets directly.
Capability inventory: Reads local video files, executes system commands (ffmpeg, ffprobe), writes frame images to the local filesystem, and optionally executes whisper for transcription.
Sanitization: The script uses os.path.basename to sanitize the video filename before creating output directories, which mitigates path traversal risks.
[EXTERNAL_DOWNLOADS]: The skill checks for the presence of the openai-whisper library and offers to install it from the official PyPI registry if missing.
Evidence: subprocess.check_call([sys.executable, "-m", "pip", "install", "--quiet", "openai-whisper"]) in scripts/understand_video.py. This uses a reputable source and includes an interactivity check (isatty) to ensure the user is prompted.
[COMMAND_EXECUTION]: The script executes ffmpeg and ffprobe as subprocesses to extract video metadata and frames.
Evidence: The script uses subprocess.run with list-based arguments (e.g., cmd = ["ffprobe", ... , video_path]) and avoids shell=True, which is the recommended safe practice for executing system commands with external input.

video-understand