How to Write AI Image Prompts That Actually Work

Most bad AI images are caused by the same thing: the prompt described a category instead of a picture.

“A businessman in an office” has a million equally valid answers, so the model hands you the average of all of them — which is exactly what a generic stock photo is. The fix is not more adjectives. It is fewer, better-chosen ones.

The structure that fixes most images

Work through five slots. You do not need all of them, but you need more than “a businessman in an office”.

Slot	Ask yourself	Example
Subject	Who or what, specifically?	a tired chef in a whites jacket
Action	What are they doing right now?	leaning on a pass, staring at a ticket rail
Setting	Where, and what is behind them?	a narrow service kitchen, steel and steam
Shot	Distance, angle, lens	medium shot, slightly low angle, 35mm
Light	Direction and quality	hard overhead fluorescents, deep shadows

Put together: “a tired chef in a whites jacket, leaning on a pass and staring at a ticket rail, narrow steel service kitchen, medium shot, slightly low angle, 35mm, hard overhead fluorescent light.”

That is not a longer prompt than most people write. It is a more specific one. Every slot removes a decision the model would otherwise make by averaging.

Why adding more words usually makes it worse

Once the picture is specified, extra adjectives start competing.

“Cinematic, dramatic, epic, moody, hyper-detailed, 8k, masterpiece” does not stack. Those words pull in overlapping directions, the model splits the difference, and you land back at the average — the exact problem you were trying to solve. The stacked-keyword style you see in prompt marketplaces is mostly cargo cult.

Change one thing at a time. Generate, look, then adjust a single slot. You will learn what your tool actually responds to in about ten minutes, which is worth more than any prompt list.

The three prompt shapes worth knowing

Describe the photograph, not the scene. “35mm, shallow depth of field, shot from across the table” tells the model where the camera is. Scene descriptions leave the camera unspecified, and an unspecified camera defaults to the boring one.

Name a medium, not a vibe. “Charcoal sketch on rough paper”, “1970s Kodachrome”, “technical line drawing” all produce a strong, consistent look. “Artistic” and “beautiful” produce nothing in particular.

Say what is behind the subject. Background is where generated images collapse into mush. One clause fixes it: “against a plain concrete wall”, “a blurred kitchen line behind”.

Pick the right tool for the job

This matters more than prompt craft for one specific case — text.

Words in the image? Use Ideogram. Readable text is its defining strength, and it is why posters, thumbnails and logo concepts go there. FLUX is also near the top of the field for text.
Beauty above all? Midjourney makes the most striking images and is the least literal about exact layouts. If you have a precise composition in mind, you will be negotiating with it.
Commercial safety inside Photoshop? Adobe Firefly, with a real caveat: Adobe’s commercial-safety claim covers its own models, and partner models inside Firefly are not covered by it. Check which model produced your image before you ship it.
Building it into your own product? FLUX is sold per image through an API with open-weight variants you can self-host, which is why it quietly powers a lot of other tools.

Our full ranking is best AI image tools, with a longer decision guide at best AI image generators, and Midjourney vs FLUX settles the most common head-to-head.

Fixing the four failures you will actually hit

Hands and faces are wrong. Crop them out or move the camera back. This is still the least reliable part of image generation, and composition is a faster fix than prompting.

Text is gibberish. Change tools rather than prompts. If you are asking Midjourney for a poster with a headline on it, no prompt will reliably rescue that.

Everything looks like a stock photo. You described a category. Add the shot and the light.

It ignored part of my prompt. Long prompts get diluted. Cut it in half and generate again — you will usually find the model was drowning in modifiers rather than ignoring you.

The habit that beats prompt lists

Keep the prompts that worked, with the tool name and the date beside them.

Image models change quickly and prompt behaviour changes with them, so a prompt that produced something excellent in March may need adjusting by August. A short personal file of what worked, for your subjects, in your tool, is worth more than any downloaded pack of a thousand prompts — because those were written for a different model version and usually a different tool.

Where to go next

Best AI image tools — our ranking, with what each one is for
Best AI image generators — job-based pick guide
Midjourney vs FLUX — artistry against control
Ideogram — the text-rendering specialist
What you actually get on free AI plans — before you pay for anything

Frequently asked questions

Why do my AI images look generic?

Almost always because the prompt describes a category rather than a picture. "A businessman in an office" has a million equally valid answers, so you get the average of them. Specify the shot — distance, angle, lens, light, what is behind the subject — and the average collapses into something particular.

Do longer prompts produce better images?

No, and this is the most common mistake. Past a point, extra adjectives compete with each other and the model splits the difference, which pulls the result back toward average. Short and specific beats long and vague every time. Add one constraint at a time and look at what changed.

Which AI image generator handles text in images best?

Ideogram is the strongest at rendering readable text, which is why it is the usual pick for posters, thumbnails and logo concepts. FLUX is also near the top of the field for text. Midjourney is the most beautiful and the least literal, so for anything with words in it you are fighting the tool.

Can I use AI images commercially?

It depends entirely on the tool and the plan, and you must check rather than assume. Adobe Firefly is built around this claim — its own models are trained on Adobe Stock and licensed content — but Adobe is explicit that partner models inside Firefly are not covered by the same commercial-safety promise. Read the terms for the specific model you used.

What is a negative prompt?

A list of things you do not want. It is useful for recurring failures — extra fingers, watermarks, text where you wanted none — but it is a blunt instrument. If a subject keeps appearing wrong, describing what you *do* want more precisely usually works better than banning what you do not.

Should I use a free image generator?

For learning, yes. Ideogram has free credits, and Adobe Firefly gives 25 free generative credits a month. Be aware that free-plan images on Ideogram are public in its community gallery, which rules it out for anything confidential.