Start with the subject, not the adjectives
Most weak prompts open with mood words. A model has nothing to anchor to until it knows what it is drawing, so the first clause should name the subject and its state: what it is, what it is doing, and where it sits in the frame.
Secondary subjects come next, and only the ones that carry meaning. A reflection, a second figure or a light source that shapes the scene is worth naming. Background clutter usually is not.
Describe composition before you describe beauty
Composition is the part of an image that survives every model. Say whether the frame is horizontal or vertical, where the subject sits, and how the eye travels. Terms like 'low three-quarter framing' or 'centred with generous negative space' are far more actionable than 'beautiful composition'.
Perspective belongs here too. Camera height changes an image more than almost any other single variable, and it costs three words to specify.
Light is the highest-leverage sentence
Name the quality, the direction and the temperature. Soft or hard, where it comes from, and whether it reads warm or cool. Those three facts reproduce a look more reliably than a list of style references.
If you cannot tell where the light is coming from, say what you can observe instead — 'even, shadowless' is honest and useful. Inventing a lighting setup you cannot see is how prompts drift away from the reference.
Only claim camera detail you can actually infer
Focal length can be estimated from perspective compression and background separation, but it is an estimate. Phrase it that way. A prompt that says 'roughly an 85mm equivalent with shallow depth of field' is credible; one that invents an aperture and a shutter speed is noise.
Finish with style and mood, then stop
Style is a compression of everything above, so it goes last and stays short. One medium, one genre, one mood. Long trailing lists of artist names dilute the concrete description you already wrote.
Imprompta follows this same order automatically: it extracts a structured reading of the image first, then composes the prompt from that structure, so the language stays grounded in what is actually visible.