How do you write a good AI image prompt?

Why the wording of a prompt matters

An image model does not read your prompt like a person. A text encoder turns the words into numbers, and the image generator is trained to make pictures that match those numbers. Words that often appeared next to certain kinds of images in the training captions pull the result toward those images. The word golden hour, for example, shows up beside warm, low-angle sunlight so often that it works almost like a lighting setting.

Two consequences follow. First, concrete visual words beat abstract ones: wet cobblestone street gives the model something to draw, while a sad mood leaves it guessing. Second, the model weighs every word, so filler such as very, beautiful or amazing takes space without adding information. Most tools also have a limit on how much text the encoder reads; words past that limit are cut or matter less, so the first part of the prompt carries the most weight.

Many generators also use guidance: they compare an image made with your prompt to one made without it and push the result in the direction of the difference. Higher guidance follows the words more strictly but can look harsh or oversaturated. This is why a clear prompt with moderate guidance usually looks better than a vague prompt with guidance turned up.

A reusable prompt formula

Most good prompts answer the same six questions. Write them in this order, because the model tends to treat early words as more important.

[subject + key details], [action or pose], [setting + time],
[style or medium], [lighting], [camera / lens / framing], [color + mood]

1. Subject: who or what, with two or three details that matter (age range, clothing, material, color).
2. Action: what the subject is doing. A verb gives the model a pose to draw.
3. Setting: where and when. One place and one time of day is enough.
4. Style or medium: photo, watercolor, 3D render, ink sketch, flat vector, anime cel shading.
5. Lighting: the single strongest lever on mood. Name the light source and its direction.
6. Camera and composition: shot size, angle, lens, depth of field, where the subject sits in the frame.

You do not need all six every time. A product shot may skip the action; an abstract pattern may skip the camera. But checking the list before you press generate catches the most common gap, which is a prompt that describes the subject and nothing else.

Weak to strong: a worked rewrite

Here is a typical first attempt. It gives the model a subject and an adjective, and nothing about how the picture should look.

a beautiful cat in a cafe, high quality, 8k, masterpiece

The result is usually a generic, centered cat with a blurry interior. 8k and masterpiece do not add detail on most current models; they only nudge toward a polished stock look. The rewrite below keeps the same idea but answers the six questions.

an orange tabby cat with a white chest, sitting on a wooden windowsill
inside a small cafe, looking out at the rain, late afternoon,
35mm film photograph, soft window light from the left, warm interior lamps
in the background, medium close-up at eye level, shallow depth of field,
muted warm colors, quiet and cozy

If the second version is still not right, change one part at a time and keep the rest fixed. Changing five things at once makes it impossible to tell which word caused the change.

Keywords for style, lighting, lens and composition

These terms are widely understood by image models because they are common in photo captions, art descriptions and film language. Use one or two from each row, not all of them.

Negative prompts: what they do and when to use them

A negative prompt is a second text field that tells the model what to move away from. In tools based on Stable Diffusion or FLUX it is usually a separate box or a negative_prompt parameter; Midjourney uses a --no parameter. Some chat-style image tools have no negative field at all, and there you should write the request positively (a clean, empty wall instead of no text on the wall).

Negative prompts work best for specific, visible things:

Prompt:   a clean product photo of a ceramic mug on a white table
Negative: text, logo, watermark, extra handles, cluttered background

• Good targets: text and watermarks, extra fingers, a second person, a specific object you keep getting, a color you do not want.
• Weak targets: long lists of quality words such as ugly, bad, worst quality. They help a little on some older models, but they also push every image toward the same look.
• Do not put an idea in both fields. A cat in the prompt and cat in the negative cancel out and confuse the result.
• Writing no inside the main prompt often backfires: a room with no windows mentions windows, and the model may draw them.

Common mistakes and honest limits

• Too many subjects. Three people with three outfits and three actions often get mixed: clothes swap, hands merge. Generate the main subject first and add others with editing or inpainting.
• Asking for exact text. Many models still misspell words or invent letters, especially in non-Latin scripts. Keep text short, put it in quotes, or add it later in an editor.
• Hands, counts and positions. Five apples or the cup to the left of the book is not always respected. Check the result and regenerate or edit instead of adding more words.
• Conflicting style words. photorealistic watercolor anime pulls in three directions. Pick one medium.
• Copying a prompt between models. A prompt tuned for one tool can behave differently in another, because each uses a different text encoder and training data. Test a few runs before you rely on it.

Non-English prompts work in many tools, but some models were trained mostly on English captions. For example, cv.cm's AI image generator translates a non-English prompt to English before it runs the model, so short, literal phrases translate more reliably than idioms.