The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes system binaries including ffmpeg and ffprobe for media processing, and uv for environment management. These commands are executed using argument lists, preventing shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: During initialization, the skill downloads the Whisper transcription model from HuggingFace and Python dependencies from official or well-known mirrors (e.g., PyTorch and Tsinghua University mirrors). These sources are recognized as safe for this type of application.
[PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it processes audio from user-provided video files into a text transcript. This transcript is subsequently analyzed by the agent, potentially allowing instructions embedded in the audio to influence agent output.
Ingestion points: Video files are processed in scripts/video-analyzer/transcriber.py to extract audio for transcription.
Boundary markers: No markers are used to delimit the transcribed content.
Capability inventory: Subprocess execution of ffmpeg and ffprobe in scripts/video-analyzer/extractor.py and scripts/video-analyzer/transcriber.py; file writing for temporary frames and audio files.
Sanitization: No sanitization is performed on the output transcript before it is returned to the agent.

video-analyze