Prompts that work on Nano Banana Pro: the four-pass shape

Part 7 of 11 · ← previous · all parts · next → Upscale or regenerate

A prompt that works on Nano Banana Pro is not a list of adjectives. It is four passes over a picture in a fixed order — what is in the frame, where the light comes from, what colour the thing is, what it is made of — and then the medium. That is the shape every decode on this site writes in, and it is the shape the model reads best. Here are three real ones, and what happened when we changed them.

The four-pass reading

A decode reads a picture the way a person who has to rebuild it would: composition, light, colour, materials. That order is not decoration. Each pass constrains the next — you cannot describe the colour of a surface until you know where the light is hitting it — and a prompt written in that order gives the model its constraints in the order it can use them.

A real one, from this series

This is the prompt a decode wrote for one of our own generations, unedited:

Atmospheric editorial photograph of a traditional brass oil lamp with a single flickering warm flame, encircled by a fresh marigold garland of bright orange and yellow blossoms, resting atop a dark polished granite plinth outside an ancient stone temple courtyard. Low-angle eye-level medium close-up shot on 85mm lens with shallow depth of field, soft cool ambient morning blue hour lighting contrasting with the warm golden flame, textured aged stone architecture, crisp specular highlights, refined focus.

Read the order. Medium, then subject, then place, then framing and lens, then light, then materials. There is not one word about mood, not one "beautiful", "stunning" or "masterpiece", and no camera brand. Every clause removes possibilities.

The same shape, a different subject

Here is a second decode-written prompt, from the clay pot in part 4:

Atmospheric editorial photograph of a rustic terracotta clay pot resting beside a folded coarse linen cloth on a weathered wooden bench, situated outside an ancient wet stone structure during misty rain. Eye-level medium close-up with shallow depth of field, softly blurring the damp masonry background. Diffuse overcast daylight illuminates cool slate-blue and earthy brown tones, capturing glistening water droplets, textured earthenware, and damp mossy stone in tack-sharp detail.

Identical architecture, nothing in common in the subject. That is what makes it a shape rather than a phrase — you can pour any picture into it.

What we changed, and what changed in the picture

In part 5 we ran a decode prompt that included the words "pulsating bioluminescent circuitry" and "monochromatic palette with selective saturation". The source picture had white sparkle on the skin and blue only in the irises. Nano Banana Pro gave us neon blue circuit traces all over the face.

The model was right and the prompt was wrong. "Circuitry" is a thing, and a thing described gets drawn. Three corrections that generalise:

• Name the behaviour, not the noun. "White specular sparkle catching on the cheekbones" beats "bioluminescent circuitry", because the first describes light and the second describes an object the model will now install.
• Say where the light falls, not how it feels. "Rim light from behind left" survives; "ethereal glow" becomes whatever the model likes.
• Put the medium first and leave it alone. "Atmospheric editorial photograph" at the front does more work than any amount of "photorealistic, 8k, ultra detailed" at the end.

Things that do nothing

From running a lot of these: quality words at the end of a prompt ("8k", "ultra-high definition", "masterpiece") are noise — the resolution setting decides the resolution, and the model does not try less hard when you do not flatter it. Negative lists are weak on this model; if you do not want people in the frame, describing an empty place works better than writing "no people". Camera bodies do little; focal length does something real, because it implies a distance and a compression.

Where to get prompts already written this way

There are more than sixteen thousand of them in the gallery, each attached to the picture it describes, each with its palette and style DNA beside it. Reading them is free. That is the fastest way to learn this shape — not by being told it, but by reading two hundred prompts that all have it and seeing the skeleton show through. Part 11 is about working that library.

FAQ

How do I write a good prompt for Nano Banana Pro?

In four passes, in order: composition, light, colour, materials — then say what kind of picture it is. Each pass constrains the next, and the model uses them in that order. Skip mood words and quality words; they do nothing.

Do quality words like 8k and ultra detailed help?

No. The resolution setting decides the resolution, and the model does not work harder for flattery. Focal length does help, because it implies a distance and a compression; camera brands do not.

Why did my prompt add something I did not want?

Because you named it as a thing. Any noun in a prompt is a thing the model may draw — "bioluminescent circuitry" got us neon traces across a face that should have had white sparkle. Describe the behaviour of the light instead of naming an object.

Do negative prompts work on Nano Banana Pro?

Weakly. Describing the scene you do want is more reliable than listing what you do not — an empty street beats "no people".

Where can I find prompts that are already written well?

The aikizi gallery holds over sixteen thousand published decodes, each carrying the prompt for the picture beside it, with its palette and style DNA. Reading them costs nothing.

Take any prompt from the gallery, change one clause, and run it in the Studio — one changed clause at a time is how you learn what each pass is doing.

Part 7 of 11 · ← previous · all parts · next → Upscale or regenerate