How to Write AI Image Prompts That Actually Work

What goes into a good prompt for generating an image?

A good prompt covers six things: subject, scene, style, light, composition and format. Every one you leave out is one the model decides for you, and models tend to decide badly: generic backgrounds, stock-photo colours, postcard people.

Think of the prompt as a brief to a photographer who has never heard of your brand. "A photo of coffee" gets you an average photo of coffee. Say which cup, on what table, at what time of day and for which network, and you get something you can use.

A short template that fits almost everything:

Subject: what appears and what it is doing.
Scene: where it is, what surrounds it, what is NOT there.
Style: photograph, flat illustration, watercolour, 3D mock-up...
Light and colour: soft window light, late-afternoon backlight, warm or muted palette.
Composition: wide shot or close detail, point of view, where the empty space sits.
Format: vertical, square or landscape, depending on where it will be posted.

What is the difference between a weak prompt and a specific one?

The specific one is checkable: you can look at the image and say whether it delivers. Here is an example with an artisan bakery called "Slow Oven".

Weak: "Nice bread photo for Instagram."

Specific: "Photograph of a freshly cut sourdough loaf with an open crumb and one slice leaning in front. Light wooden table, grey linen cloth, a dusting of flour. Soft natural window light from the side, early morning, gentle shadows. Slightly tilted overhead shot, empty space at the top. Warm, muted palette, no people, no logos. Vertical 4:5 format."

The second one has subject, scene, light, composition, palette and format. It also says what it does not want, which matters more than it looks.

Why tell the model what you do NOT want?

Because it removes the decorations the model adds out of habit. Many models fill a scene with props: candles, flowers, steam, hands, invented text. One line of exclusions prevents that: "no text, no logos, no hands, no people, no watermarks".

Two practical rules:

Put exclusions in a short list at the end, not tangled into the description.
Some models have a separate negative-prompt field, others respond better to positive wording ("plain background" instead of "no objects in the background"). Test both with your model.

How do you iterate without losing what already works?

Change one variable per attempt. If you alter light, style and framing at once, you cannot tell what improved the image or what broke it.

A method that works:

Generate a first version with the full prompt.
Write down what is wrong. Usually it is one thing: cold light, a cluttered background, a crop that is too tight.
Rewrite only that sentence and generate again.
When a version is close, stop rewriting and edit it with words: "remove the glass in the background", "make the light warmer", "leave more space at the top". Editing an image that is nearly right usually beats generating a new one from scratch.

Keep the prompts that work next to the brand for next month. A good prompt is an asset: swap the subject, keep the light, style and palette, and the images look like the same series.

Can you ask for text inside the image?

You can ask, but do not count on it coming out right. Image models still often fail at writing: warped letters, words with an extra character, inconsistent type. It varies from model to model and changes over time, so test yours.

A safer approach:

Generate the image without text, leaving clear space in the frame (that is why "empty space at the top" or "on the left" is worth asking for).
Add the headline afterwards in your usual editor, in your brand typeface.
If you really need text inside the image, keep it to one or two short words and check every letter.

A misspelled name, price or address on a published image is a mistake everyone sees.

How do you adapt the prompt to each publishing format?

Ask for the format from the start instead of cropping later. An image composed as vertical and cropped to square often loses the subject or leaves the air in the wrong place. We cover which ratio each network uses in image and video formats for social media.

When writing the prompt, include:

The ratio (vertical, square, landscape).
Where the subject should sit so the app interface does not cover it (the top and bottom edges of vertical formats are often covered by buttons and captions).
How much space to leave free if you will overlay a headline later.

What does an image cost, and how do you avoid surprises?

It costs whatever the model you pick costs, and the sensible move is to see that before generating. In igroui, every generation is paid in credits, the estimated cost is shown before you launch it, and the final charge is the real cost, which can vary slightly. If the generation fails, the credit reservation is released.

To spend less without losing quality:

Test the idea on a cheaper model and keep the pricier one for the final version.
Do not ask for ten variations of something you have not defined: sharpen the prompt first.
Reread the prompt before launching. Most wasted attempts are vague prompts.

If you want to see how this fits with your brand, the AI image generator feature describes generating and editing from text, and brand voice explains how to keep a consistent identity across pieces.

What should you check before publishing an AI-made image?

Check five things: details, text, resemblance to others, rights and transparency.

Details: hands, teeth, reflections, objects melting into each other, shadows that do not add up. Zoom in and scan the whole image.
Text: every letter, if there is any.
Resemblance to others: it should not echo a recognisable brand, artwork or person. If in doubt, regenerate.
Rights: the terms of use of the model you used and of the platform you publish on. If you upload your own images as reference, make sure you have the right to use them.
Transparency: tell your audience the image was made with AI. That is the recommended practice and, depending on the network and the country, may be mandatory. Whoever generates and publishes is responsible for the result, not the tool.

One last rule of thumb: if the image tries to pass as a real photo of a product or an event that does not exist, do not publish it. Use AI images to illustrate ideas, moods and concepts, not to prove something that never happened.