How to write AI image prompts

Start with the image you want to describe

An image prompt describes the result you want a model to attempt. Start with a clear subject and action, then add the setting, composition, lighting and material details that matter. A prompt recipe keeps this text together with the model, settings and references where they are available. Specific visual choices are easier to evaluate than a list of quality adjectives.

Build a prompt from concrete visual decisions

Use this original example as a starting point. Replace the subject and setting to fit your idea. The camera position and lighting describe observable choices; they do not guarantee that a model will follow every instruction.

A small ceramic teapot on a pale oak table beside a window. Close-up at table height, teapot placed on the left third with empty space on the right. Soft morning side light, gentle shadows, matte glaze with subtle handmade texture. Quiet cream and sage palette, editorial product photograph.

Keep model settings separate from the prose

When the tool supports them, record the model version, aspect ratio, seed and reference images in their respective settings. Do not assume that a command or negative-prompt syntax from one model works in another. If a setting is missing from a shared work, treat it as unknown rather than inventing a value. In Aveniqa, open a work and inspect its prompt recipe and creation details.

Change one thing and compare

For a first revision, keep the subject, composition and model fixed and change only the lighting, for example from morning side light to soft overcast light. Compare the outputs for that specific difference. Record the change with the result so another person can understand what you tested. Even the same prompt can produce different outputs across model versions and generation runs.

Use shared works as inspectable examples

Browse the prompt library, choose an image, and read the prompt alongside its result. Check the source and rights information before reusing a reference asset. Public visibility does not by itself grant permission to reuse every image. Copying a prompt is a useful starting point for study, but reproducing a work may also depend on its references, model settings and later edits.