The Agent Skills Directory

[COMMAND_EXECUTION]: The Python scripts (generate_video.py, prep_video.py, inspect_video.py, and upload_file.py) use the subprocess module to interface with ffmpeg and ffprobe for media processing and metadata extraction.
Evidence: Use of subprocess.run() in multiple scripts to perform video trimming, scaling, and inspection.
Mitigation: The skill uses list-based command construction rather than shell strings, and explicitly checks for the presence of the required binaries before execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text prompts and video files which are transmitted to the Gemini API for generation.
Ingestion points: The prompt argument in generate_video.py and input file content.
Boundary markers: The skill structures the prompt with specific tags (e.g., <FIRST_FRAME>, <VIDEO_REF_0>) to help the model distinguish between instructions and reference data.
Capability inventory: File writing to the local media/ directory and network communication with Google APIs via the official SDK.
Sanitization: The skill implements a robust sanitize_error function that redacts API keys, OAuth tokens, and internal infrastructure URLs from error logs, reducing the risk of accidental credential exposure.
[EXTERNAL_DOWNLOADS]: The skill downloads generated video assets from Google's official API endpoints.
Evidence: The download_video_file function in generate_video.py uses the google-genai SDK to fetch content from generativelanguage.googleapis.com.
Context: These downloads target a well-known service and are a core requirement for the skill's primary function.

gemini-omni-flash-api