How to Write Better Prompts for Nano Banana 4

Most disappointing AI images are caused by vague prompts, not model limitations. This guide gives you a repeatable structure that turns "meh" generations into images you'd actually use.

The four-part prompt structure

Build prompts in this order:

1. Subject — who or what, with specific details. Not "a woman" but "a woman in her 60s with silver hair and a linen shirt".
2. Setting & action — where the subject is and what's happening. "Standing at a market stall, picking up a peach."
3. Light & mood — the emotional temperature. "Late afternoon sun, warm and unhurried."
4. Style & technical specs — the rendering instructions. "35mm documentary photography, shallow depth of field, natural colors."

Put together:

A woman in her 60s with silver hair and a linen shirt,
standing at a market stall picking up a peach,
late afternoon sun, warm and unhurried atmosphere,
35mm documentary photography, shallow depth of field, natural colors

Each part answers a question the model would otherwise have to guess.

Speak camera, not adjectives

"Beautiful" and "high quality" do almost nothing. Camera language does:

You don't need to be a photographer — steal these phrases and the model does the rest.

Name a style anchor

The single highest-leverage trick: anchor your image to a recognizable visual tradition.

• "national geographic photography style"
• "Kodak Portra 400 film look"
• "studio ghibli background art"
• "1970s sci-fi book cover illustration"
• "architectural digest interior photography"

One good anchor outperforms ten scattered adjectives.

Say what you don't want — positively

Nano Banana 4 responds better to positive instructions than to "no X". Instead of "no text, no watermark", write "clean image with no text". Instead of "don't make it cartoonish", write "photorealistic with natural skin texture". Frame every constraint as a description of the desired result.

Iterate in small steps

When an image is close but not right, change one thing per message:

• "Same image, but make the lighting warmer"
• "Keep everything, replace the background with a beach at dusk"
• "Zoom out to show the full body"

The model keeps the rest stable. Changing three things at once makes it hard to know what caused what.

A checklist before you hit enter

• Does the prompt name a specific subject with at least two concrete details?
• Is there a light source or time of day?
• Is there exactly one style anchor?
• Did you specify aspect ratio if it matters?
• Is it under ~80 words?

Print this list, use it five times, and you won't need it anymore. Then explore the prompt library to see the structure applied across portraits, products and scenes.