How to write better AI image prompts

Most people writing prompts for AI image generators do one of two things: they write a single sentence and hope for the best, or they paste a 400-word stack of style keywords copied from a Discord server. Both produce inconsistent results. There's a better, simpler middle ground — and once you internalize it, your hit rate goes up dramatically.

This guide walks through a six-part structure that works reliably with Google's Nano Banana 2 and Nano Banana Pro models. It's the same structure we use internally when testing the models for NoHiddenPrompt.

1. Lead with the subject

Start your prompt with what the image is actually about. The first ten words carry the most weight. Don't bury the subject behind a list of style modifiers. 'A bronze samurai helmet on a velvet cushion' tells the model exactly what to render. 'Cinematic, ultra-detailed, 8k, photorealistic, a helmet' tells it nothing for the first five tokens.

2. Describe the action or pose

If your subject is a person, animal, or object that could be doing something, say what it's doing. 'A woman reading' is weak. 'A woman leaning against a window, holding a paperback book at chest height, eyes down' is strong. Specific actions ground the image and force the model to commit to a composition.

3. Set the environment

Where is this happening? A 'red brick wall covered in moss' is a far more interesting backdrop than 'outdoors.' One concrete environment beats five vague atmospheric words. Pick a location and describe two or three details that anchor it.

4. Add lighting

Lighting is the difference between a snapshot and a photograph. You don't need cinematography vocabulary — plain English works. 'Hard sunlight from the left, deep shadows on her face' or 'soft overcast light with no shadows' are both clearer than 'dramatic lighting.'

5. Specify the medium and style — last

Save style for the end. 'Shot on 35mm film, slight grain' or 'oil painting on canvas, visible brushstrokes' or 'flat vector illustration, two-color palette.' Style instructions at the end of the prompt act as a finishing layer rather than overwhelming the subject.

6. Cut anything that doesn't earn its place

Read your prompt out loud. Every word should add information the model can act on. 'Beautiful' and 'amazing' don't. 'Trending on Artstation' is meaningless to Gemini models. 'High quality' is implied. Trim ruthlessly.

A worked example

Bad: 'A samurai, ultra detailed, 8k, masterpiece, dramatic lighting, trending on artstation, photorealistic, high quality, beautiful'

Good: 'A bronze-armored samurai standing in a narrow Kyoto alley at dusk, hand resting on the hilt of a katana, soft warm light spilling from a paper lantern beside him, light rain darkening the cobblestones, shot on Kodak Portra 400, slight film grain.'

The good version is longer, but every clause carries weight. The model has a subject (samurai), action (standing, hand on hilt), environment (Kyoto alley, cobblestones, lantern), lighting (soft warm dusk light, rain), and style (Portra 400 film). The bad version is just adjectives.

Iterate, don't restart

When you get a result that's 80% right, don't throw the prompt away. Use the 'Use as reference' button in NoHiddenPrompt to feed the result back as a reference image, then make a small change to your prompt — adjust the lighting, swap the location, change the pose. The model will preserve what was working and update what you asked it to.

Common mistakes to avoid

• Stacking five lighting descriptions ('golden hour, cinematic, dramatic, moody, volumetric'). Pick one.
• Negating things ('no people, no text'). Models are inconsistent with negatives. Describe what you want instead.
• Asking for very specific text in the image. All current image models are unreliable at rendering text — keep it short or skip it.
• Specifying camera lenses without context ('85mm f/1.4'). It works sometimes but is wildly inconsistent. Describe the look you want instead.

Prompting is a skill, but it's not a mystical one. Subject, action, environment, lighting, style. Trim the rest. Iterate on what's close. That's the whole game.