Nano Banana Prompt Maker

First, the name. Google announced Gemini 2.5 Flash Image in August 2025 and said in the same post that it goes by “nano-banana”. This GPT writes prompts aimed at that kind of model. It is mine, not Google’s, and I have no affiliation with Google or with OpenAI beyond publishing a GPT in their directory.

Second, the important part: it produces text. It does not generate images. You take what it writes and run it in whatever tool you use.

Open Nano Banana Prompt Maker

What it does

Most weak image prompts fail for one of two reasons. They leave out the thing the model needed most, or they ask for two incompatible things at once: studio lighting and available light, a wide establishing shot and legible small text, minimalism and seven objects.

So this GPT does two jobs. It asks for what’s missing: what the image is for, what has to stay fixed, what must not appear, aspect ratio, where it will be seen. Then it reads the request back and names the contradictions before writing anything.

Output is usually a short prompt and a longer one, plus a list of what it deliberately left unspecified so you know what the model is free to invent.

The place it saves the most time is edits. Writing an instruction that changes the background while keeping the product identical is harder than writing a prompt from scratch, and this is the workflow I built it for.

What it’s bad at

Knowing whether a prompt worked. It never sees the image. It cannot learn from your results within a conversation unless you describe what came back, and “describe what came back” is a skill in itself. Tell it what was wrong in specific terms and the second attempt is usually much better. Tell it “bad” and you’ll go in circles.

Model-specific syntax. Weights, negative prompts, seeds and the parameter flags of any particular tool change often. It doesn’t reliably track them. Its prompts are plain descriptive language, which travels between tools but leaves tool-specific control on the table.

Text inside images, hands, precise counts and exact brand colour. No prompt phrasing reliably fixes these. If it implies otherwise, it’s overstating what a prompt can do.

Taste. It can eliminate contradictions. It cannot tell you an image is good. That judgement stays with you or with a designer, and the gap between a technically clean prompt and a picture worth publishing is exactly where the work is.

Faces of real people, copyrighted characters, or a competitor’s protected brand assets. It won’t help with those and neither will I. That’s the provider’s policy and it’s also the correct answer.

Who shouldn’t bother

If you already write image prompts fluently, this is overhead. You’ll write a better prompt directly than you will by describing your idea to something else first.

If you need one image once, just describe it to your image tool. The structure here pays off across a set, whether that is variants, campaign consistency, or repeated edits to a single product shot.

If your visual direction isn’t decided yet, no prompt tool helps. Decide the direction with references, then come back.

Things worth asking it

• “Here’s my idea in one messy paragraph. Before writing anything, tell me what you’d need to know and what in this contradicts itself.”
• “One concise prompt and one detailed prompt. Then list what you left unspecified.”
• “Keep this product exactly as it is. Change only the environment. Write the edit instruction.”
• “This prompt gave me an image that was too soft and too symmetrical. Fix the prompt, and tell me which change addresses which problem.”
• “Six variants that differ on one variable at a time, and name the variable for each.”

The last one matters more than it looks. If a variant set changes five things at once, you learn nothing from picking the winner.

What it’s for

Ad and landing-page visuals, mostly edits rather than fresh generations. It is the narrowest of the three GPTs listed here, and narrow is the point: it does one part of the job (turning an idea into an unambiguous instruction) and hands everything else back to you.