AI Image Prompts: How to Write Them So They Work

Most disappointing AI images come from prompts that describe a thing instead of a picture. "A woman drinking coffee" is a thing. A photograph has a lens, a light source, a mood, a distance from the subject and a reason the frame is cropped where it is.

You don't need a secret word list. You need to say the five things a photographer or illustrator would have decided before pressing the shutter, then iterate on one at a time.

The five-part structure for AI image prompts

Build every prompt from these, in roughly this order:

1. Subject. Who or what, and what they're doing. Be specific: "a chai vendor pouring from height into a glass" beats "a man with tea".
2. Setting. Where, and what's in the background. This is what stops the image looking like a stock photo on a white backdrop.
3. Style or medium. Photograph, oil painting, 3D render, line illustration, risograph print. If you want a photo, say photograph, and name a lens: "35mm", "85mm portrait", "wide angle".
4. Lighting and mood. Golden hour, overcast, harsh midday sun, single lamp, neon reflections on wet street. Lighting changes an image more than almost any other word.
5. Composition. Close-up, full body, from above, shallow depth of field, centred, rule of thirds.

Put together:

"A chai vendor pouring tea from height into a glass, busy Mumbai street stall at dawn, documentary photograph, 35mm, warm low side light, steam catching the light, shallow depth of field, close-up on hands"

That prompt makes decisions. The first version, "a man making tea", leaves them all to the model.

Parameters: where tools differ

Parameters are the switches that control output rather than describing it, and they're tool-specific.

Midjourney. Its parameter documentation describes parameters as "special instructions you can use to guide how your images turn out" and sets two rules: "Always put parameters after your prompt text" and "Don't use commas, periods, or other punctuation marks in your parameters." The docs show vibrant California poppies --ar 2:3 as correct, and flag three mistakes: no space before the dashes, punctuation in the parameter, and prompt text placed after the parameters.

The two you'll use constantly are --ar for aspect ratio (images start square, so --ar 16:9 for a banner, --ar 9:16 for a phone screen) and --no to say what you don't want in the image.

Ideogram. The one to reach for when the image contains words. Put the exact text in quotes and keep it short; long sentences inside an image still break.

Adobe Firefly. Fewer cryptic switches, more on-screen controls for aspect ratio, content type and style. Good when you want commercially careful imagery without learning a parameter syntax. See our Adobe Firefly vs Canva comparison.

Leonardo. Strong on consistent styles and reusable characters across a set of images.

Iterate on AI image prompts one variable at a time

This is the habit that separates people who get good images from people who regenerate forty times:

1. Write the prompt and generate four.
2. Pick the closest one, even if it's wrong in three ways.
3. Change exactly one element: the lighting, or the lens, or the time of day.
4. Generate again and compare.

Changing five words at once tells you nothing about which word mattered. Keep a note of prompts that worked; your own list beats any downloaded prompt pack, because it's tuned to what you make.

Describe what you want, not what you don't

Models respond better to positive description. "Empty street at dawn" works better than "a street with no people", because the words you use get represented either way. When you genuinely need to exclude something, that's what Midjourney's --no parameter is for.

The same applies to quality words. Stacking "8k, ultra detailed, masterpiece, award-winning, trending on artstation" was folklore from early models and mostly wastes prompt space now. Describing the actual light does more.

Prompting for specific jobs

• Thumbnails: leave empty space on one side for text, ask for high contrast, and keep the subject big. Canva is the easiest place to add that text afterwards.
• Logos: generate concepts, then rebuild the final mark yourself. Tool licences and copyright rules make a raw generated logo risky.
• Product shots: name the surface and the background ("on brushed concrete, soft overhead light, plain grey background").
• People: describe role, age range and expression rather than naming a real person. Many tools block real-person prompts, and generating a recognisable public figure causes problems you don't want.

Things that still break

• Hands and crowds are better than they were, but complex interactions between multiple people still go wrong.
• Text, except in tools built for it, and even then check every letter.
• Specific brands, logos and copyrighted characters, which most tools block or render badly.
• Exact counts. "Five birds" often gives you four or seven. Crop instead of fighting it.

For the tools themselves, see our best AI image generators ranking, the free AI image tools list, and DALL·E vs Midjourney.

FAQ

How do I write a good AI image prompt?

Name the subject, the setting, the style or medium, the lighting and the composition, then add tool parameters. Generate four images, change one element, and generate again.

Where do parameters go in a Midjourney prompt?

At the end, after the prompt text, with no punctuation inside them. Midjourney's docs give vibrant California poppies --ar 2:3 as the correct form and warn against commas, missing spaces, or text after the parameters.

How do I set the aspect ratio?

In Midjourney, use --ar at the end of the prompt, such as --ar 16:9 for widescreen or --ar 9:16 for vertical. Images default to square. Other tools offer aspect ratio as an on-screen setting.

Which AI tool handles text inside images best?

Ideogram is the one known for rendering short words legibly. Keep the text short, put it in quotes, and check every letter before using the image.

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

Rarely with current models. Describing the lighting, lens and composition does far more than stacking quality adjectives.