How I Use Claude to Write My Nano Banana Image Prompts

I don't write my own image prompts. I haven't for a long time. I describe what I want to Claude, let it write three versions of the prompt, and paste those into Higgsfield. That's the whole method, and it's the reason the images I use actually match what I had in my head instead of being whatever the generator felt like producing that day.

Most people do the opposite. They open an image generator, type "image for my AI course website", and take what comes out. Then they wonder why it looks generic. It looks generic because the prompt was generic. The fix isn't learning prompt syntax. The fix is getting something that's good at language to write the prompt for you.

Start with the purpose, not the picture

Here's the request I actually made, roughly as I said it: I want an image for BotBrained.ai that showcases the learning experience between teacher and student. A teacher demonstrating AI on their computer, for YouTube viewers.

Read that back. It's vague. I don't know how to quantify "the learning experience" either. That's fine — that's the point. I'm not trying to write a prompt. I'm describing an outcome and letting Claude figure out how to turn an abstract idea into something a model can render.

Then I add the part that does the work:

Give me three unique prompts to create an image in Nano Banana.

Three, not one. Unique, not variations. That single word is why this works — instead of one prompt and three near-identical outputs, you get three genuinely different takes on the same idea. Different angles, different scenes, different framing. You pick.

I was using Claude through the chat harness for this, on Opus 5. Worth knowing: Opus 5 takes longer. Around thirty seconds before it started producing. If you're used to instant responses, don't assume it's broken.

What comes back

The prompts are verbose. Much more detailed than anything I'd sit down and type by hand. They spell out the scene, the lighting, the perspective, the mood, the aspect ratio. That detail is the whole value — it's the part I'd skip if I were doing this myself, and it's the part that makes the difference between a usable image and a stock-photo-looking miss.

One of the three came back specifying 16:9. So that's the aspect ratio I used for it. The prompt knew what shape it wanted to be.

Getting them into Higgsfield

Higgsfield is where I do all my image and video generation. Straightforward once you know the layout: click Image, then Create Image. Model selector is at the bottom — I'm on Nano Banana Pro. Aspect ratio is bottom left.

Paste the prompt, set the ratio, generate. Repeat for the other two.

Now, the thing nobody tells you.

The unlimited plan trap

I have an unlimited plan. Higgsfield does not respect it by default.

Every single time I generate, the interface is set to charge me tokens. I have to manually switch unlimited on before I hit generate. If I don't notice, that's two tokens gone on an image I'm already paying a subscription to make.

I'll say it plainly: that's shady. It's not a bug, it's a default, and the default costs you money. Check the toggle every time.

Two more things about unlimited:

It caps at 2K. If you want 4K, you pay tokens. I use 2K for everything because it's free and I don't need better than that for what I'm doing.

You get queued. Unlimited generations go to the back of the line behind everyone paying with tokens. I proved this in the video without meaning to — I generated three on unlimited, then generated one with tokens several seconds later. The token one finished first, while the other three were still queued.

So there's a real decision to make. If I'm in a rush, I turn unlimited off, pay the two tokens, and get it now. If I'm not, I wait and it costs nothing. Knowing the tradeoff exists is the useful part.

The results

All three came back close to what I asked for. One had somebody on YouTube teaching AI workflows to a student viewer — exactly the brief. One was in-person with a video camera, coach pointing at the screen. One was the same idea from a completely different angle and perspective.

That's what asking for three unique prompts buys you. Claude didn't give me the same prompt three times with the adjectives shuffled. It said: here's this angle, here's this scene, here's this perspective.

And the one I paid two tokens for? Basically identical to one of the free ones. It wouldn't have mattered. I could have just made all three on unlimited and waited.

Controlling the style

You don't have to accept photorealistic. Say what you want:

• "I don't want these photorealistic, I want them more stylized"
• "Give me this in Pixar style"
• "3D rendered"
• Game art, editorial, whatever fits

You explain the output you're looking for and Claude writes the prompt that gets there.

If you don't know what your options are, ask. Literally: what different styles can I create images in? You'll get the breakdown — 3D, artistic, stylized, black and white, the lot.

One detail from that answer that's worth keeping. Nano Banana responds to style as a described look, not a keyword tag. So the phrases that reliably land are things like editorial, documentary photography, cinematic film still, commercial product photography. You're describing a kind of photograph, not tagging a genre. That distinction changes your results more than any other single thing.

The part I haven't set up yet

I have the Higgsfield MCP connected to my Claude Code. In principle I can skip the copy-paste entirely — tell Claude to generate three images like this, it writes the prompts, runs them through Higgsfield, and hands the images back.

I haven't wired that into this workflow yet. That's a future video.

One warning I'll give now, though: I'm fairly sure the MCP burns tokens rather than using the unlimited plan. It behaves more like API access than like the web interface. So automating it isn't automatically free.

Where this goes

The same approach scales past single images. I've used it to plan whole movie trailers and book trailers — tell Claude I want a one-minute film trailer, feed it the source material and how things should look, and it plots the thing out. Complete workflow, a diagram, then: here are all the images you need for the opening scenes, here are the prompts for each. Or with the MCP running, it just goes and makes them.

Different scale, same move. Don't write the prompt. Describe the result and let the model that's good at language handle the language.

If you want to see the clicking rather than read about it, the video walks through the whole thing start to finish. And if you're building something with this and want people to react to it, that's what BotBrained.ai is for — post what you made and find out what other people think.