Ten Common Prompt Mistakes and What to Write Instead

Most failed generations come down to a small set of repeating mistakes. Here they are together — with the reason and the replacement.

1. Evaluative words instead of description

"Beautiful", "professional", "masterpiece", "high quality". These carry no visual information: the model does not know what you find beautiful, and it already aims at plausibility.

Replacement: concrete attributes — light, shot size, palette, optics.

2. Negations

"Without a hat", "not blurred", "no text". The concept enters the description regardless of the negation.

Replacement: a positive description of what should be in that place.

3. Too many requirements

A request with twenty conditions is followed worse than one with six: each requirement carries less influence.

Replacement: seven or eight meaningful elements maximum, priorities first.

4. No light

The commonest omission. Without it you get even, sourceless illumination — the primary tell of the AI look.

Replacement: direction, character and temperature of the light in three words.

5. Contradictions

"Soft dramatic contrasty light", "a minimalist detailed scene". The model picks one, unpredictably.

Replacement: check the request for compatibility before sending.

6. Camera settings instead of visual outcomes

Long lists of exposure, aperture and ISO barely affect the result: the model does not simulate a camera.

Replacement: the visual consequences of those settings — "shallow depth of field", "motion blur", "grain".

7. Expecting precision in small things

"Exactly three buttons", "the sign reads…", "the clock shows 10:15". The model neither counts nor writes.

Replacement: remove these elements from the frame or add them later in an editor.

8. Describing appearance instead of uploading a photo

"A woman of thirty, dark hair, green eyes" specifies a type, not a specific person.

Replacement: photo-based generation when a recognisable person is required.

9. Rewriting the request instead of rerunning

The spread between runs is often larger than the effect of an edit. Rewriting after the first failure loses what was working.

Replacement: two or three runs unchanged, and only then an edit — one element at a time.

10. Ignoring the format

Aspect ratio determines composition. Cropping a finished frame leaves a fragment of a composition designed for something else.

Replacement: set the format before generating, based on the platform.

How to check yourself

A useful habit: before sending, read the request and ask whether it lets you picture one specific image. If several different pictures come to mind, the model will pick one of them too — and not necessarily yours.

A second check: are there words that add nothing? Usually, after removing them, the request becomes both shorter and more precise.