The Agent Skills Directory

[SAFE]: The skill manages credentials securely by instructing the user to store the Gemini API key in a dedicated environment file (~/.nanobanana.env), which is loaded using the python-dotenv library. This is a recommended practice to avoid hardcoding secrets.
[SAFE]: All network operations are conducted through the official google-genai library to communicate with Google's API infrastructure. Google is a recognized trusted organization, and no unauthorized external domains were detected.
[SAFE]: Input images are processed using the standard Pillow (PIL) library, and file system interactions are restricted to reading specified inputs and writing outputs to user-defined paths. The script includes logic to ensure output directories exist before writing.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a standard attack surface for AI agents by processing user-supplied text and images that are passed to an external model.
Ingestion points: Text prompts and image file paths are accepted via the --prompt and --input CLI arguments in nanobanana.py.
Boundary markers: The skill relies on the structured API request format provided by the Google GenAI SDK to distinguish between instructions and data; no additional local delimiters are applied to the prompt strings.
Capability inventory: The script is capable of reading local image files, writing images and metadata to the disk, and performing network requests to the Gemini API.
Sanitization: Inputs are handled by robust standard libraries (Pillow for images, JSON for metadata), but the text prompt itself is passed to the model without local sanitization, relying on the model provider's safety filters.

nanobanana-skill