How to Write AI Image Prompts That Actually Work

Most AI image prompts fail for the same handful of reasons, and almost none of them are about the model you picked. After running a few hundred generations across Gemini and ChatGPT, the pattern is consistent: the prompts that work are specific in the places that matter and quiet everywhere else.

Why most prompts come out wrong

A vague prompt forces the model to guess, and it will guess toward the average of everything it has seen. "A man in a nice shirt standing outside" describes ten million photographs. You get the blandest one.

The fix is not a longer prompt. It is a prompt that is precise about the four things the model cannot infer.

Specify the subject, the light, the framing, and the finish. Leave everything else alone — over-specifying colour and mood tends to fight the model rather than steer it.

The four things worth spelling out

1. The subject, including what must not change

If you are working from a reference photo, say so explicitly. The single highest-impact line you can add is a face-lock instruction at the very start of the prompt, before anything else.

• Say preserve facial identity 100% before you describe anything else
• Name the build and posture — "naturally lean, relaxed upright posture"
• Describe the expression in plain words, not adjectives like "beautiful"

2. The light

Light is what makes an image read as a photograph instead of a render. Name the source and its direction.

• "natural diffused daylight" for soft, believable skin
• "harsh overhead fluorescent" for a cold, editorial look
• "bright vertical panel directly behind the subject" for a rim-lit profile

3. The framing

Camera language works because the training data is full of it. A focal length and an aperture will do more for you than three sentences of description.

You want	Add this
Full body, slightly heroic	35mm, slightly low camera angle
Tight portrait, soft background	85mm at f/1.8
Instagram / Pinterest crop	vertical 3:4 or 2:3 aspect ratio

4. The finish

End the prompt with the texture you want, not the emotion. "Visible natural skin pores, realistic fabric texture, photorealistic RAW DSLR quality" gets you further than "stunning, gorgeous, masterpiece".

Run it more than once

This is the tip people skip. The first generation is almost always the softest one. Run the same prompt two or three times without changing a word — the second and third results are consistently sharper in skin texture and fabric detail.

A quick checklist before you hit generate

1. Is the face-lock line the first thing in the prompt?
2. Have you named the light source and its direction?
3. Is there a focal length, an aperture, or an aspect ratio?
4. Did you name actual fabrics instead of writing "shirt"?
5. Is your reference photo sharp, well lit and unobstructed?

Get those five right and the model stops guessing. That is the whole trick.