agricidaniel/banana-claude ? reverse-engineered prompt

Build me a Claude Code plugin that lets me create, edit, compare, and review Gemini images from simple plain language prompts.

I want it to feel like a safe creative assistant that plans the image request first, shows me the prompt, model choice, settings, output location, and rough cost estimate before anything paid happens, then waits for my approval. After that it should generate the image, save the result plus a small metadata file, and let me inspect the output so I can regenerate or tweak it if needed.

It should support making new images, making one clear edit to an existing image, continuing an earlier session, comparing a few prompt and model options, and adding final text or logos locally when exact wording matters. Keep the flow offline first where possible, and make sure keys and private info are never exposed in logs or output. If you need current Gemini or Claude Code details, look them up online first.