How to write AI image prompts that actually work

By Ali Ghafori · 3 min read

AI image generators are powerful, but vague prompts give vague results. "A cool car" will get you a car — just not the one you imagined. Good prompts aren't about magic keywords; they're about describing the picture the way a photographer or art director would. Here's the structure I use.

The five building blocks

Almost every strong prompt answers five questions, roughly in this order:

1. Subject — what is in the picture? Be specific: not "a woman" but "a young woman in a yellow raincoat holding a paper map".
2. Setting — where is it? "On a narrow cobblestone street in an old European town, after rain."
3. Style — what kind of image is it? A photograph, a 3D render, a watercolor, an anime frame, a product shot?
4. Composition — how is it framed? Close-up, wide shot, low angle, centered, rule of thirds, shallow depth of field.
5. Light and mood — golden hour, soft overcast light, neon at night, dramatic side light, calm, energetic, mysterious.

Putting it together

Weak:   a cool car in the city

Strong: a vintage red sports car parked on a wet city street at night,
        neon signs reflecting on the asphalt, low-angle wide shot,
        cinematic photograph, shallow depth of field, moody blue and
        magenta lighting

The second prompt leaves far less to chance. Every phrase removes a decision the model would otherwise make randomly.

Use photography language

Image models were trained on huge numbers of captioned photos, so camera and lighting terms work very well:

• Shot type: extreme close-up, portrait, medium shot, full body, wide establishing shot, aerial view.
• Lens feel: 35mm (natural), 85mm (flattering portraits), wide-angle (dramatic perspective), macro (tiny details).
• Depth of field: "shallow depth of field" or "bokeh background" to isolate the subject.
• Lighting: golden hour, softbox studio lighting, rim light, backlit, overcast, high-key, low-key.

Iterate one thing at a time

When a result is close but not right, don't rewrite the whole prompt. Change one element — the lighting, or the angle — and generate again. If you change five things at once, you won't know which change helped. Keep a simple text file of prompts that worked; over time it becomes your personal style library.

Keeping images consistent

Consistency is the hardest part, especially when you need the same character or product across many images, like a thumbnail series or a storyboard.

• Reuse a fixed description block. Write the character once ("a man in his 30s with short curly black hair, round glasses, green bomber jacket") and paste it into every prompt unchanged.
• Use reference images when your tool supports them. Image-to-image and character reference features keep faces and outfits far more stable than text alone.
• Lock the style words too — the same lighting and style phrases across a series make the set look intentional.

From image to video

Many AI video tools now animate a still image. This is often the most controllable way to make AI video: get the image exactly right first, then describe only the motion in the video prompt.

Image prompt: a lighthouse on a rocky cliff at sunset, dramatic clouds,
              cinematic wide shot

Video prompt: slow drone push-in toward the lighthouse, waves crashing on
              the rocks, clouds drifting, light beam starting to rotate

Keep motion prompts simple and physical: camera movement plus one or two things that move in the scene. Asking for too much action in a few seconds usually produces warping and artifacts.

Finishing touches

Generated images are rarely final. I usually upscale the best result, fix small problems (hands, text, odd edges) with inpainting or an image editor, and color-correct it so it matches the rest of a project. For video, AI clips get the same treatment as normal footage: trimmed, graded, and given sound design in the editor.

← Back to all articles