The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the ffmpeg command-line utility via Python's subprocess.run to extract audio from video and sample video frames. These calls use list-based arguments, which prevents shell injection, and are necessary for the skill's media processing functionality (found in audio_features.py and video_features.py).
[EXTERNAL_DOWNLOADS]: The skill interacts with the Gemini API to upload media files and retrieve semantic analysis results. It uses the official google-genai library and handles authentication through environment variables. These interactions target a well-known service (Google) for their intended purpose.
[DATA_EXFILTRATION]: Media files and metadata from user-provided spreadsheets are sent to the Gemini API. This is the primary function of the Gemini pipelines and is documented in the references. The data is sent to a trusted service provider.
[PROMPT_INJECTION]: The skill processes untrusted metadata from spreadsheet columns and interpolates it into prompts for the Gemini model (format_metadata in utils.py). While this creates a surface for indirect prompt injection, it is managed by the system prompt and the intended use case for semantic labeling.
[DYNAMIC_EXECUTION]: The Gemini pipeline scripts dynamically load a configuration file (pipeline_config.py) using importlib. This is used to define project-specific logic like metadata formatting and validation rules, which is a standard pattern for this type of complex data pipeline.

q-multimodal