The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data by accepting up to 20 image URLs via the image_urls parameter. As acknowledged in the skill's security section, these images are fetched by the remote model server and represent a surface for image-based prompt injection or adversarial content designed to influence the model's behavior.
Ingestion points: image_urls array in the google/nano-banana-2/edit endpoint defined in SKILL.md.
Boundary markers: The skill documentation includes a dedicated "Security & Privacy" section that warns users to treat external URLs as untrusted and identifies image-based prompt injection as a known risk.
Capability inventory: The skill invokes the runcomfy CLI to perform operations and writes generated output files to a user-defined local directory via the --output-dir argument.
Sanitization: The documentation states the CLI transmits the JSON body directly to the API without shell-expanding the prompt, providing protection against command injection, though it does not perform content validation on the source images themselves.
[COMMAND_EXECUTION]: The skill is designed to invoke the @runcomfy/cli tool using the runcomfy run command. It constructs a command-line call that passes user-supplied prompts and image references as a JSON payload to the google/nano-banana-2/edit endpoint.
[EXTERNAL_DOWNLOADS]: The skill instructions require the installation of the @runcomfy/cli package from the npm registry. Additionally, during execution, the tool fetches generated image assets from the *.runcomfy.net and *.runcomfy.com domains to save them to the local filesystem.

nano-banana-edit