How to write AI art prompts that actually work

The five things that actually change a result

Most weak prompts fail for the same reason: they name a subject and stop. The model then has to invent everything else, and it will invent something generic. A prompt that produces sellable work usually specifies five things.

1. Subject, with one specific detail

"A lighthouse" gives you a stock lighthouse. "A squat granite lighthouse with a rust-streaked lantern room" gives you something that looks observed rather than imagined. One concrete detail does more work than three adjectives.

2. Medium

This is the single biggest lever on style. Name the physical process you want the image to imitate: oil on canvas, loose watercolour washes, charcoal on toned paper, linocut print, gouache, 35mm photograph. Without it, most models default to a glossy digital render, which is the hardest look to sell as art.

3. Light

Light is what makes an image feel like a place rather than a diagram. Try golden hour backlight, overcast north light, single candle, deep shadow, harsh midday sun, blue hour. Naming the light also fixes the mood, so you often do not need a separate mood word at all.

4. Composition

Tell the model where to stand: low horizon, sky filling two thirds, tight crop on the hands, wide establishing shot, flat-on, centred, symmetrical. This matters more than usual for prints, because a composition that works at thumbnail size may look empty at 24×36.

5. Palette

Two or three colours, named. Muted ochre, sage and off-white. Cold greys with one red accent. A stated palette is what makes a series of prints look like a collection rather than a pile.

A worked example

What does not help

• Stacking quality words. "Masterpiece, 8k, ultra detailed, award winning, trending" adds noise, not quality. Modern models do not need it.
• Very long lists. Past roughly sixty words most models start weighting your later terms less. Cut rather than add.
• Contradictions. "Minimalist and highly detailed" makes the model pick one at random. Decide which you mean.
• Negative instructions. "No people" often produces people. Describe what should be there instead: "empty street at dawn".

Iterating without burning credits

Changing five things at once tells you nothing about which one worked. A cheaper, faster loop:

1. Draft the prompt with all five elements present.
2. Run it on a low-cost model — Stable Diffusion Core or FLUX Schnell at one or two credits an image — purely to test the composition and subject.
3. Change one element. Run again. Compare.
4. Once the prompt is right, re-run it on your final model. That is the only expensive generation you need to pay for.

The More like this button re-runs a prompt without retyping it, and every prompt is saved under Saved prompts with the image it made, so a phrasing that worked in March is easy to find in September.

Using Get inspired

If you are starting cold, the Get inspired button takes a couple of keywords, optionally an artist and an art movement from over a hundred options, and drafts a full prompt in the structure above. Treat the result as a first draft — swap the light or the palette for something of your own and it stops looking like everyone else's output.

Prompting for a series

Collections sell better than singles. The trick is to hold four of the five elements constant and vary only the subject:

[SUBJECT VARIES], loose watercolour washes on rough paper,
late afternoon backlight, low horizon with sky filling
the upper two thirds, muted ochre, sage and warm grey.

Run that with "sunflower field", "wheat stubble", "flooded meadow" and "hedgerow in October" and you have four prints that clearly belong together — which is what makes someone buy the set rather than one.

Next

Once the prompt is doing its job, the model choice is the other half of the result. See choosing the right AI image model, and making AI art that prints well for the things that only go wrong on paper.