How to Write a Prompt for an AI Image

If you need a prompt for an AI image right now, use this shape:

[subject], [what it's doing], [where], [style], [lighting], [framing]

Worked example: a border collie asleep on a windowsill, small apartment kitchen, warm morning light, shallow depth of field, shot on 35mm film.

That's it. Six slots, and you can leave any of them out. The rest of this covers why each slot matters, what to change when the result is close but wrong, and the specific things image models handle badly so you don't waste attempts on them.

The vocabulary below is shared across the major image models — Stable Diffusion, Midjourney, and the image tools built into general assistants all respond to the same descriptive slots, which is why prompt habits transfer between them.

The one rule that matters most: describe what you want to see, not what you don't. Image models have no reliable mechanism for negation in the prompt itself — "no people in the background" frequently produces people. Describe the empty street instead.

The Six Slots, and What Each Controls

If you want to see how varied the open ecosystem is, Hugging Face hosts thousands of image models, and the same six slots work across nearly all of them.

Style and lighting do the most work. Two prompts with identical subjects and different style words produce completely different images; two with different subjects and the same style look like siblings. If you're unhappy with a result and don't know why, change the style word first.

Be specific in the right places

Specificity helps on nouns and materials, and hurts when you over-specify things you don't care about.

• Good: "brass door handle, worn at the edges" — concrete, visual
• Wasteful: "a beautiful, stunning, amazing, high-quality masterpiece" — these are not visual descriptions and mostly consume attention
• Counterproductive: fifteen adjectives on the subject and none on the setting, which produces a detailed object floating in a generic space

Quality-booster words ("8k", "masterpiece", "trending") were folklore from earlier models and matter far less on current ones. Spend the words on the image instead.

Fixing a Result That's Close

The most useful skill here isn't writing the first prompt — it's diagnosing the second.

Right subject, wrong mood → change lighting and setting, leave the subject alone.

Right mood, wrong subject → the subject description is competing with the style. Simplify the style, make the subject more concrete.

Too busy → remove elements rather than adding "simple". Prompts describing three things produce three things; prompts describing one produce one.

Composition is wrong → this is framing, and it's the slot people leave out. "Close-up", "wide shot", "from above", "centred" are instructions the model follows reasonably well.

Everything is slightly generic → you've written categories rather than specifics. "A dog" is a category. "A wet terrier shaking itself dry" is an image.

Change one slot at a time. Rewriting the whole prompt when one element is wrong means you learn nothing about which word did what, which is the same discipline that applies to text prompts — and the same reason model providers' prompting guides push testing one variable at a time.

What Image Models Still Do Badly

Knowing these saves more time than any prompt trick, because you stop attributing them to your wording.

Text in images. Improving, still unreliable, and worse the longer the string. Wikipedia's prompt engineering entry covers how text-to-image conditioning differs from text generation, which is the underlying reason. For anything where the text must be exact, generate the image and add the text yourself.

Precise counts. "Five birds" produces roughly five birds. Small numbers are more reliable than large ones; exact counts are not something to depend on.

Hands and complex anatomy. Much better than it was, still where errors cluster. Poses that hide hands are a legitimate workaround.

Spatial relationships between multiple objects. "The red cup to the left of the blue book, behind a lamp" degrades quickly with each added relation. Two objects is usually fine; four is a gamble.

Consistent characters across images. Getting the same person in two generations is hard from prompting alone, and it's what character-reference and image-conditioning features exist to solve.

None of these are prompt problems, which is the point — no rewording fixes them. Design around them.

Style Without Copying a Living Artist

A note worth taking seriously: naming a living artist to copy their style is legally and ethically contested, and several platforms restrict it.

You can get the same result by describing what you actually want, which is usually more precise anyway. Instead of an artist's name, specify:

• Medium — oil, gouache, linocut, 3D render, charcoal
• Era or movement — art nouveau, bauhaus, 1950s advertising
• Technique — cross-hatching, flat colour, heavy impasto
• Reference format — technical manual diagram, polaroid, screen print

This is better prompting regardless of the ethics, because "in the style of [artist]" is a single fuzzy token and "linocut, three flat colours, heavy black outlines" is four precise instructions.

Working the Other Direction

If you've seen an image whose look you want and can't describe it, you can go backwards — tools that analyse an image and produce a text description of it. Our guide to turning an image into a prompt covers the free options and their limits.

This is genuinely the fastest way to learn style vocabulary. Run five images you admire through one, and the recurring words are the vocabulary you were missing.

If your bottleneck is less about wording and more about getting a working image setup running at all, Taku mirrors AI apps and workflows into a desktop workspace and runs them without the environment setup — the free app library is where to look. Taku is in Beta, and the Mac app is available now.

Key Points

• Six slots: subject, action, setting, style, lighting, framing. Leave any out; don't leave out style
• Describe what you want, not what you don't — negation in prompts is unreliable
• Style and lighting are the biggest levers. Change those first when a result feels wrong
• Quality-booster words are largely obsolete — spend the words describing the image
• Change one slot at a time, or you learn nothing from the attempt
• Text, exact counts, hands, and multi-object spatial relations are model limits, not prompt failures
• Describe medium, era, and technique rather than naming a living artist — more precise and less contested

FAQ

What is a good prompt for an AI image?

One that names the subject concretely, says what it's doing and where, and specifies style, lighting, and framing. For example: "a border collie asleep on a windowsill, small apartment kitchen, warm morning light, shallow depth of field, shot on 35mm film." Concrete nouns and a clear style do most of the work.

What are the best prompts for AI art?

There's no universal list, because the best prompt depends on the look you want. What transfers is the structure — subject, action, setting, style, lighting, framing — plus specific medium and technique words instead of vague quality adjectives. Collect style vocabulary from images you admire rather than copying prompt lists.

Why does my AI image not match my prompt?

Usually one of three things: you described what you didn't want (models handle negation poorly), you wrote categories rather than specifics, or you hit a known model limitation like exact counts, readable text, or complex spatial relationships. Change one element at a time to find out which.

Do quality words like "8k" and "masterpiece" help?

Much less than they used to. They were useful on earlier models and are largely superstition on current ones. Words spent on medium, lighting, and framing do more than words spent on generic quality claims.

How do I get a consistent character across several images?

Prompting alone is unreliable for this. Use whichever image-reference or character-consistency feature your tool provides, since that's the problem those features exist to solve. Detailed text descriptions of a face will still drift between generations.