The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input (text prompts and image data) which is passed to the Gemini model for processing. This is a standard risk for LLM-based tools.
Ingestion points: User-supplied instructions and image file paths are ingested in scripts/generate_image.py, scripts/edit_image.py, scripts/compose_images.py, and the library scripts/gemini_images.py.
Boundary markers: None. Input instructions and image objects are passed directly to the google-genai client without explicit delimiters or system instructions to ignore embedded commands.
Capability inventory: The skill can perform network operations (connecting to the Gemini API) and file system operations (reading source images and saving generated content to disk using PIL.Image.save()).
Sanitization: No sanitization or validation of the input text instructions is performed before they are sent to the model.
[EXTERNAL_DOWNLOADS]: The skill downloads dependencies (google-genai and Pillow) from official package registries. These are well-known, trusted libraries for the intended functionality.
[COMMAND_EXECUTION]: The skill provides CLI scripts that execute logic locally for image processing. These scripts use standard argument parsing and do not involve shell injection vulnerabilities.

gemini-imagegen