How to Write AI Image Prompts That Work

Type “a cool picture of a city” into an AI image generator and you’ll get something. It will probably be shiny, generic and forgettable. The model filled every gap you left with its most average guess.

Good image prompts work the way a good brief works for a photographer or illustrator. You describe what’s in the frame, how it looks and how it should feel. This guide gives you a simple formula and shows how to fix the most common problems.

The six-part image prompt formula

Not every prompt needs all six parts, but going through them makes sure you’ve thought about each one.

1. Subject: what’s in the picture

Be specific about the main thing and what it’s doing.

• Weak: a dog
• Better: a scruffy grey terrier sitting on a windowsill, looking outside

2. Setting: where it is

Give the background and surroundings.

…in a small apartment kitchen, rain on the window, plants on the counter

3. Style: what kind of image

This changes the result more than anything else. Name a medium or a look:

• photograph, 35mm film photo, product photo on white background
• watercolor illustration, flat vector illustration, pencil sketch
• 3D render, isometric illustration, pixel art

Describe a style instead of naming a living artist. “Loose watercolor with soft edges and muted colors” gives you the look without copying anyone.

4. Lighting: how it’s lit

Lighting sets the mood fast.

• soft morning light, golden hour, overcast daylight
• dramatic side lighting, neon glow, candlelight
• bright, even studio lighting

5. Composition: how it’s framed

Tell the model where the “camera” is.

• close-up, wide shot, overhead view, eye level
• subject on the left, empty space on the right (useful if you’ll add text later)
• shallow depth of field, blurred background

6. Format and mood

Finish with the aspect ratio and feeling.

• 16:9 for a blog header, 1:1 for social media, 9:16 for stories
• calm, playful, tense, nostalgic

Many tools have a separate aspect ratio setting. Use it if it’s there, because it’s usually more reliable than writing the ratio in the prompt.

Putting it together

Before:

a person working on a laptop

After:

A young woman working on a laptop at a wooden desk by a large window,
small home office with bookshelves and a plant. Natural photograph,
soft afternoon light from the left, shallow depth of field. Medium shot
from slightly above, subject on the left third, empty wall space on the
right. Calm, focused mood. 16:9.

The second prompt gives you a usable blog header, with room on the right for a title.

A template you can copy

[Subject and action], [setting]. [Style/medium], [lighting].
[Shot type and camera angle], [composition notes]. [Mood]. [Aspect ratio].

Fill in the brackets, delete the parts you don’t care about, and generate.

How to fix common problems

It looks too generic. Add specific, unusual details: textures, colors, objects in the background. “A ceramic mug with a chipped rim” beats “a mug.”

The style keeps changing. Put the style near the start and repeat key words like flat vector illustration, minimal, two colors. When you find a result you like, save the exact prompt and reuse it for a consistent series.

Text in the image is wrong. Keep it to a few words and put them in quotation marks: a sign that says “OPEN”. If it’s still wrong, add the text afterwards in a design tool. That’s usually faster and always accurate.

Hands, faces or small details look odd. Newer models are much better at these, but they can still slip. Regenerate, use the tool’s edit or inpainting feature on just that area, or pick a composition where those details are less prominent.

Too much is happening. Long lists of objects confuse models. Cut the prompt down to one main subject and two or three supporting details.

Change one thing at a time

When a result is close but not right, don’t rewrite the whole prompt. Change one element and generate again:

• Same prompt, different lighting
• Same prompt, different camera angle
• Same prompt, different style

This shows you what each word is doing, and you’ll get better results faster. If the tool lets you set a “seed” or start from a previous image, use it to keep everything else stable while you adjust.

Use a chat assistant to draft prompts

If you’re stuck, a text assistant can turn a rough idea into a detailed prompt:

I need a blog header image about learning to code with AI. Write three
detailed image prompts in different styles (photograph, flat illustration,
3D render). Include subject, setting, lighting, composition with space for
a title on the right, mood and a 16:9 ratio.

Then paste the best one into your image tool. The same principles from our six-part text prompt framework apply: specific beats vague.

Use images responsibly

• Check the tool’s terms for commercial use, especially on free plans.
• Don’t generate real people in misleading situations, or trademarked characters and logos.
• Label AI images where readers would reasonably expect a real photo, such as news or product reviews.
• Add alt text describing the image when you publish it. It helps visually impaired readers and search engines.

Key takeaways

• Build prompts from six parts: subject, setting, style, lighting, composition, and format/mood.
• Style and lighting change the result the most, so always include them.
• Describe styles in your own words instead of naming living artists.
• Fix problems by changing one element at a time, and save prompts that work.
• Add short text and fine details afterwards when the model struggles.

Want ready-made text prompts too? See our prompt templates for beginners.