The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of text prompts and image files which are passed to a multimodal AI model without explicit boundary markers or sanitization logic.
Ingestion points: The args.prompt and args.images arguments in generate_image.py accept external data.
Boundary markers: Absent; the script directly appends user input to the model's content list.
Capability inventory: The skill possesses Write access to save images and Bash access to execute the generator script.
Sanitization: None detected; raw input is passed to the Google GenAI client.
[EXTERNAL_DOWNLOADS]: The skill depends on the google-genai Python package.
Evidence: The generate_image.py script specifies google-genai in its inline dependency metadata.
Context: The package is a well-known library from a trusted organization for interacting with official AI APIs.
[METADATA_POISONING]: The skill description and usage instructions reference a model version (gemini-3-pro-image-preview) that is currently not part of the public Google Gemini model lineup.
Evidence: Found in SKILL.md description and generate_image.py model parameter.
Context: While likely a placeholder or speculative name, it qualifies as deceptive metadata regarding the skill's actual backend capabilities.

nano-banana-pro