Writing Better Image Prompts

Why short prompts disappoint

When you write two or three words, the model fills every unstated decision with the statistical average of its training data. That is why minimal prompts produce images that feel stock: centred subject, even lighting, neutral background, mid-distance framing. Nothing is wrong with any individual choice, but together they read as generic because they are the most common option in every category at once.

The fix is not to write longer prompts for their own sake. It is to notice which decisions you actually care about and state those, letting the model default on the rest. A prompt that specifies lighting and composition precisely while saying nothing about colour is usually better than one that vaguely gestures at everything.

A structure that works

Most reliable image prompts cover four things in roughly this order: what the subject is, how it is framed, how it is lit, and what visual tradition it belongs to. Working through those in order stops you from over-describing the subject and forgetting everything else, which is the most common failure.

• Subject: who or what, and what they are doing. Be concrete about the action.
• Composition: distance, angle and what fills the frame. Close-up, wide shot, low angle, overhead.
• Lighting: direction, quality and time of day. Soft window light, hard midday sun, single rim light.
• Style: the visual tradition. Documentary photography, technical illustration, oil painting, screen print.

Be specific about light

Lighting does more to determine whether an image feels intentional than any other single factor, and it is the thing people most often leave out. Saying an image is lit by a single source from the left, or by overcast daylight, or by a screen in a dark room, changes the result far more than adding another adjective to the subject.

Describing light also implicitly fixes mood, contrast and colour temperature without you having to name them, which is why it is such an efficient use of prompt space.

Name a tradition, not a person

Style is best specified by naming a medium, process or era rather than a living artist. Technical illustration, mid-century travel poster, large-format landscape photography, ink and wash all carry precise visual information about line, colour and composition.

Naming a living artist to imitate their style raises real ethical and legal questions, and many platforms restrict it. Describing what you actually want visually is both more reliable and avoids the problem entirely, because it tells the model about the image rather than about a person.

Iterate on one variable

When a result is close but wrong, change one thing. If you rewrite the whole prompt you learn nothing about which change helped. Adjust the lighting alone, or the framing alone, and keep the version that improves.

Keep prompts that work. Most people end up with a handful of reliable scaffolds they adapt rather than writing from scratch each time, and building that library is most of the skill.

• Change one category per iteration so you can attribute the difference.
• Save prompts that worked, along with the model that produced them.
• If a subject keeps coming out wrong, describe it more concretely rather than adding style words.
• Generate several variations before judging a prompt, since sampling variance is large.

Negative space in the prompt

It is worth deliberately leaving some things unspecified. If you do not care about the background, saying nothing gives the model freedom to produce something coherent, whereas a half-hearted background description often produces a distracting one. Prompting well is as much about deciding what not to say as what to include.