AI Tools11 min read

How to Write AI Art Prompts: Structure, Modifiers and 18 Examples

Most bad AI images come from prompts that describe a subject and nothing else. Here is the seven part structure that fixes it, the modifiers that reliably change an image, the four things image models still cannot do, and a copyable example for every common use case.

Type a beautiful landscape into any image generator and you will get something competent, generic and immediately recognisable as AI output. Type six carefully chosen clauses instead and the same model produces something you would actually use.

Nothing about the model changed. What changed is how much of the decision making you handed over. Every detail you leave out is a detail the model fills with its statistical default, and those defaults are what make AI images look like AI images.

The Seven Parts of a Working Prompt

Think of a prompt as a shot list rather than a description. Each of these parts answers a question the model would otherwise answer for you.

1. Subject. Who or what, with the specific details that matter. Not a woman but a woman in her sixties with close cropped grey hair and reading glasses.

2. Action or pose. What they are doing, or how the object is oriented. Static prompts produce static, catalogue looking images.

3. Setting. Where, and when. A kitchen at 7am with steam on the window is a scene. A kitchen is a set.

4. Composition and shot. Close up, medium shot, wide establishing shot, overhead flat lay, low angle, shot from behind. This is the single most underused control and often the one that transforms an image.

5. Lighting. Soft window light, hard midday sun, golden hour backlight, single overhead lamp, neon spill, overcast diffuse. Lighting determines mood more than colour does.

6. Style or medium. Oil painting, 35mm film photograph, flat vector illustration, watercolour, 3D render, charcoal sketch. Be specific about medium before reaching for artist names.

7. Colour and mood. A limited palette instruction such as muted earth tones or high contrast teal and amber does more work than any adjective about atmosphere.

You will not need all seven every time. You should know which ones you are deliberately leaving out.

The same idea at three levels of detail

AttemptPromptResult
WeakA cat in a houseCentred cat, flat light, generic room
BetterAn orange tabby cat sleeping on a windowsill in a sunlit roomRecognisable scene, still a stock photo feel
StrongClose up of an orange tabby cat asleep on a wooden windowsill, warm late afternoon sun through dusty glass, shallow depth of field, 35mm film photograph, muted warm paletteSpecific image with intent

The third prompt is not longer for the sake of it. Every clause removed a decision from the model.

Modifiers That Reliably Change the Output

Some terms consistently move an image. Others are decoration that people copy from prompt lists without noticing they do nothing.

CategoryTerms that workNotes
Shotclose up, medium shot, wide shot, overhead, low angle, dutch angleStrongest single lever
Lens35mm, 85mm portrait, macro, wide angle, shallow depth of field, bokehOnly meaningful for photographic styles
Lightgolden hour, backlit, rim lighting, softbox, harsh flash, candlelit, overcastSecond strongest lever
Mediumoil on canvas, gouache, linocut, 3D render, pencil sketch, screen printChanges the image more than any style word
Era1970s Kodachrome, Victorian engraving, 1990s camcorder, Bauhaus posterCarries palette and texture at once
Palettemonochrome, limited palette, pastel, high contrast, desaturatedPrevents the default oversaturated look

Two terms worth retiring. Masterpiece and highly detailed were useful signals to older Stable Diffusion checkpoints trained on tagged aesthetic scores. On modern models they mostly add nothing and can push output towards an overprocessed style. Trending on ArtStation belongs to the same era.

Artist names are a separate matter. They are powerful shortcuts, but naming a living artist to imitate their style is contested ethically and increasingly restricted by the tools themselves. Naming the technique rather than the person gets you most of the way there and does not build your work on someone else's name.

Four Things Image Models Still Get Wrong

Knowing the failure modes saves more time than any prompt trick.

Negation. Writing without a hat frequently produces a hat, because the concept is in the prompt either way. Only a dedicated negative prompt field handles exclusion reliably. Otherwise, describe the positive alternative.

Counting. Ask for exactly five apples and you will get four, six or seven. Small counts up to about three are usually fine. Beyond that, generate and check rather than trusting the number.

Text. Better than it was, still unreliable past a couple of short words. Plan to add real typography yourself.

Precise spatial relationships. A red cube to the left of a blue sphere and behind a green cone is the kind of instruction models scramble. Keep spatial requirements simple, or compose multiple elements in an editor.

Hands, once the running joke of the genre, have largely been fixed in current models. Fingers still go wrong at small scale or in complex poses, which makes hands worth checking rather than worth fearing.

Iterate on One Variable at a Time

The most common mistake after a mediocre result is rewriting the entire prompt. You then have no idea which change helped.

A better loop:

1. Get a rough composition you like, ignoring style.

2. Lock the subject and setting wording. Do not touch them again.

3. Change only the lighting clause. Generate. Compare.

4. Change only the medium or style clause. Generate. Compare.

5. Adjust the palette last, since it is the least destructive change.

Save the prompts that worked. A personal library of six reliable structures beats any list of a thousand prompts you found online, because yours are tuned to the tool you actually use.

Aspect ratio is worth setting deliberately too. Most tools default to square, and a square frame quietly kills any composition that wanted to be a landscape or a portrait.

18 Prompt Starters by Use Case

Copy these, then swap the subject. The structure is the reusable part.

Use casePrompt
Profile pictureHead and shoulders portrait of a person with short dark hair, soft window light from the left, plain warm grey backdrop, 85mm lens, shallow depth of field, natural colour
Blog headerWide overhead flat lay of a wooden desk with a notebook, coffee cup and pen, soft diffuse daylight, muted earth tones, plenty of empty space on the right for text
App icon conceptFlat vector illustration of a folded paper aeroplane, bold single colour on a solid background, thick even strokes, no gradients, centred, generous margin
Product shotStudio photograph of a matte ceramic mug on a seamless light grey backdrop, single softbox from above left, soft shadow, sharp focus, neutral colour
Fantasy landscapeWide establishing shot of a valley of terraced fields under low mist, distant mountain ridge, cool blue hour light, oil painting with visible brushwork, limited palette
Character conceptFull body character sheet of a desert traveller in layered linen robes, neutral standing pose, flat lighting, plain background, ink and watercolour
Food photoClose up of a bowl of noodle soup with steam rising, dark wooden table, single warm overhead light, shallow depth of field, rich contrast
ArchitectureLow angle photograph of a brutalist concrete stairwell, hard midday sun casting sharp diagonal shadows, monochrome, high contrast
Retro poster1960s travel poster of a coastal town, flat shapes, limited four colour palette, screen printed texture, bold simplified forms
Children's bookGouache illustration of a small fox looking up at a lantern in a snowy forest, soft edges, warm limited palette, gentle mood
Cyberpunk streetMedium shot of a rain slicked alley at night, neon signs reflecting in puddles, backlit figure with umbrella, teal and magenta palette, 35mm film grain
Minimal wallpaperAbstract composition of soft overlapping gradients in deep indigo and violet, subtle grain, vertical format, no subject, plenty of negative space
Nature macroMacro photograph of a dew covered fern frond, morning light, extremely shallow depth of field, green and gold palette, sharp focus on the near edge
Portrait paintingThree quarter view portrait in the manner of Dutch golden age painting, single candle light source, dark background, warm skin tones, oil on canvas
Isometric sceneIsometric 3D render of a tiny cafe interior, soft ambient occlusion, pastel palette, clean edges, white background
Editorial illustrationConceptual illustration of a person carrying an oversized clock up a staircase, flat shapes, limited palette of ochre and slate, subtle paper texture
StickerDie cut sticker illustration of a smiling cactus, thick white outline, bold flat colours, simple shading, white background
Logo explorationSimple geometric mark combining a leaf and a droplet, single weight line art, black on white, centred, no text

Where This Runs on a Phone

You do not need a desktop workflow for any of this. Generai puts text to image generation and an AI chat assistant in one iPhone app, which is a practical combination for prompt work: you can draft and refine the prompt in chat, then generate from it without switching apps and losing the thread.

Two honest limits. Every consumer generator inherits the failure modes above, so no app will reliably spell a sentence or count seven objects, and any listing that suggests otherwise is overselling. And phone generation is tuned for speed and convenience rather than the fine grained control of a desktop pipeline with model weights and samplers, which is the right trade for most people and the wrong one if you need reproducible seeds.

If you are comparing options first, AI art generator apps for iPhone covers the landscape, and best AI chatbot apps for iPhone covers the text side.

The Bottom Line

A prompt is a set of decisions. Make them yourself, in this order: subject, action, setting, shot, lighting, medium, palette. Anything you skip, the model decides, and its defaults are the reason so much AI art looks the same.

Then change one clause at a time, keep the versions that worked, and stop expecting the model to spell, count or arrange objects in precise space. Those are your job, or your editor's. For the rest of what these tools are genuinely good at, how to use AI for productivity covers the workflows worth building.

Frequently Asked Questions

What makes a good AI art prompt?

Specificity in the right places, not length. A good prompt names the subject, what it is doing, where it is, how the shot is framed, how it is lit, and what medium or style it imitates. Missing any of those means the model picks for you, and its default choice is usually a centred subject in flat lighting. Long prompts are not automatically better. Once you pass roughly sixty words, later terms start to dilute rather than add, and contradictory instructions produce a muddled average of both.

Do negative prompts actually work?

Only in tools that offer a dedicated negative prompt field, which is mostly the Stable Diffusion family. There, terms placed in that field steer generation away from those concepts and are genuinely useful for removing artefacts. In a plain text box, writing no text or without a hat usually fails, because the model reads the concept and often includes it. The reliable fix in a single text box is to describe what you do want instead. Replace no background clutter with a plain seamless grey backdrop.

Why can AI image tools not spell words correctly?

Text rendering has improved a lot in recent models, and short words on signs or labels now often come out right, but it remains the least reliable part of image generation. The models were trained to reproduce the visual appearance of letterforms rather than to spell, so long strings, unusual fonts and multiple text elements in one image still degrade quickly. If the text matters, keep it to a couple of words, generate several variations, and expect to add the final wording yourself in an editor.

Who owns the images an AI generator creates?

It depends on the tool and your country, so read the terms of the specific service before commercial use. Most consumer generators grant you broad rights to use what you create, sometimes with restrictions on free tiers. Copyright status is a separate question. The US Copyright Office has held that purely machine generated output without meaningful human authorship cannot be registered, and other jurisdictions differ. In practice you can usually use the image, but you may not be able to stop anyone else from using something similar.

Try Generai: AI Chat & Art Creator

Mentioned in this article. Download free from the App Store.

More Articles