Skip to content
OpenDream AI

How to Write AI Art Prompts That Actually Work

The eight building blocks of a clear image prompt, how to order them, why specificity beats length, and how to iterate without losing what already works.

5 min read · Updated September 2026

What a prompt actually does

A prompt is a description of an image that does not exist yet. The model reads it and tries to produce something consistent with every part of it. Two things therefore matter more than anything else: whether each phrase describes something visible, and whether the phrases agree with each other. A prompt full of abstract praise gives the model little to work with; a prompt with conflicting instructions forces it to compromise in ways you did not choose.

The eight parts of a strong prompt

You rarely need all eight, but it helps to know what each one controls.

1. Subject

Who or what the image is about. Be concrete about kind, age, material, color and action. "A dog" is a starting point; "an elderly grey greyhound curled on a woven rug" is a subject.

2. Setting

Where and when the subject exists. Location, era, weather and time of day all belong here. Setting often carries more of the mood than the subject does.

3. Composition

How the image is framed: close-up, medium shot, wide shot, overhead, low angle, centered, rule of thirds, negative space on one side. Composition words are among the most reliable controls you have.

4. Light

Direction, quality and color of light: "soft window light from the left", "hard midday sun with short shadows", "neon signage reflecting on wet pavement". Light is the fastest way to change the emotional tone without changing the subject.

5. Lens or medium

For photographic images, use lens and camera language: 35mm, 85mm portrait lens, shallow depth of field, macro, long exposure. For illustration, name the medium: watercolor on cold-press paper, gouache, charcoal, linocut, 3D clay render.

6. Style

The broader visual language: editorial, cinematic, mid-century illustration, minimalist, surreal. Describe the qualities you admire, such as "loose ink linework with muted washes", rather than naming living artists.

7. Mood

The emotional register: serene, tense, playful, melancholic, triumphant. One or two mood words are plenty; stacking five dilutes them.

8. Constraints

Practical requirements: "no text", "plain light grey background", "leave space at the top for a title", "centered product with even margins". On models that support negative prompts, some constraints can move there instead.

Ordering: lead with what matters most

Models tend to give more weight to what appears early and what is stated clearly. A practical order is subject, setting, composition, light, medium and style, mood, then constraints. That puts the thing you care about first and the fine print last.

If a detail keeps getting ignored, move it earlier or give it its own short sentence. If the style is overpowering the subject, move the style later or soften its wording.

Specificity beats length

Long prompts are not better by default. Specific prompts are. Compare two approaches to the same idea:

Beautiful fantasy forest, magical, amazing detail, epic, stunning, best quality
An old beech forest in late autumn, a narrow path covered in copper leaves, low morning fog between the trunks, pale sun behind the trees, medium-wide shot at eye level, painterly realism with soft edges

The first contains almost no visual information: every word is a judgment, not a description. The second is similar in length, but every phrase can be drawn.

A useful test: for each phrase, ask whether a painter could act on it. If not, cut it or replace it with something concrete.

Before and after examples

Portrait.

Before: A woman, portrait, realistic, good lighting
After: Head-and-shoulders portrait of a woman in her sixties with short silver hair and a navy wool coat, standing by a rain-streaked café window, soft diffused daylight from the right, 85mm lens, shallow depth of field, warm muted color grade, quiet and reflective

The rewrite adds an age, wardrobe, setting, light direction, lens language and a mood. Each addition removes a guess the model would otherwise make.

Product.

Before: A cool perfume bottle ad
After: A faceted glass perfume bottle with amber liquid on a slab of pale travertine, single hard light from the upper left casting a long sharp shadow, seamless warm beige background, centered with generous empty space above, clean commercial studio photograph, no text

The rewrite specifies material, surface, lighting, background and layout, and explicitly rules out lettering.

Illustration.

Before: Cute city illustration
After: A small hillside town of stacked pastel houses at dusk, lit windows, a funicular climbing the slope, seen from across a bay, flat vector illustration with a limited palette of teal, coral and cream, gentle grain texture, cheerful and calm

"Cute" becomes a palette, a time of day, a viewpoint and a rendering style.

Iterating without losing ground

Treat each generation as feedback. When a result is close, change one element and keep the rest word for word. Useful adjustments:

  • Too busy: add "simple composition" or "few elements", or specify where the empty space should be.
  • Too flat: add directional light and a depth cue such as foreground framing.
  • Wrong scale: state relative size, e.g. "a tiny figure at the base of an enormous cliff".
  • Style drift: reinforce the medium in a second phrase, e.g. "watercolor, visible paper texture".
  • Unwanted objects: describe the desired alternative rather than only forbidding the object.

Keep a copy of each version so you can return to an earlier one if a change makes things worse.

Common mistakes

  • Using quality buzzwords instead of visual description.
  • Mixing incompatible media, such as "oil painting, 3D render, pencil sketch".
  • Burying the subject at the end of a long list of style terms.
  • Asking for many characters with individual details in one image; start with fewer.
  • Expecting long passages of text to render exactly; keep lettering short and use a text-capable model.

A reusable prompt skeleton

[Subject with key details], [setting and time], [shot type and angle], [light direction and quality], [lens or medium], [style descriptors], [mood], [constraints]

Fill in what you know, leave out what you do not care about, and let the rest be discovered through iteration. The goal is not a perfect first prompt but a clear one that you can improve step by step.

Keep learning

Put it into practice

Try what you just read in the Create Studio.

Start creating