Turn a Visual Brief Into Image Directions

A brief has four layers

Almost everyone writes an image prompt from the bottom of this diagram upward. They start with the feeling — cinematic, warm, editorial — and add the subject somewhere in the middle of the sentence. Then when the result is wrong, they adjust the feeling again.

Work from the top instead, because the layers fail in a strict order of severity. An image that misses the constraint layer is unusable no matter how beautiful it is. An image with the wrong subject cannot be rescued by lighting. An image with a great subject and mediocre style is a small fix.

The constraint layer is the one most often left out entirely, and it is the cheapest to state. "Premium product image" is not a brief. This is:

CONSTRAINTS   4:5 and 1:1 crops from the same frame.
              Top third must stay clear for a headline.
              No visible brand marks, no readable text, no faces.

SUBJECT       A single ceramic pour-over dripper, matte white,
              empty, sitting upright on a flat surface.

COMPOSITION   Centred, shot slightly above eye level,
              full object visible with margin on all sides.

STYLE         Soft directional light from the left, gentle shadow
              to the right, warm neutral background, no props.

That brief can be checked. Someone can look at a result and say "the top third is not clear" — a factual statement rather than a taste one. Most disagreement about generated images is actually disagreement about an unwritten constraint.

Diagnose the layer before you rewrite

This is the single habit that most shortens an image workflow.

When a generation is wrong, name the layer that failed before touching the prompt. The failure modes cluster very cleanly. If the subject is wrong, duplicated, or fused with something else, no amount of style vocabulary will fix it — you need fewer things in frame and plainer nouns. If the composition is wrong, you need explicit crop and camera language, not a mood.

The trap is that adding style adjectives always changes the image, so it always feels like progress. "Cinematic, 8k, highly detailed, masterpiece" applied to an image with the wrong subject produces a more beautiful wrong image, and it does so convincingly enough that people run the loop four or five times.

A practical rule: if you cannot say which layer the last change was aimed at, you are decorating.

One more diagnostic worth knowing. Generative models handle relationships between objects far less reliably than they handle the objects themselves. "A cup to the left of a book" fails often; "a cup" and "a book" each succeed. When a brief depends on a spatial relationship, either reduce the scene to one object, or plan to composite — and decide that before spending twenty generations discovering it.

Review on four axes

"I don't like it" is not a next action. Scoring the four axes separately produces one.

The axes are ordered by how expensive they are to fix. Subject accuracy is a prompt problem. Composition is a prompt problem with different vocabulary. Brand fit is often a specification problem — the brief was underspecified and the model filled the gap. Technical quality, though, is frequently a sampling problem rather than a language problem: the same prompt with a different seed produces clean hands where the last one produced a mess.

That last distinction saves a great deal of time. When the only fault is a rendering artifact, do not rewrite. Regenerate. Writing "perfect hands, five fingers" is folklore; sampling three more times is mechanism.

Generate three variations per brief rather than one. A single output tells you nothing about whether a flaw came from your brief or from the sample — three tells you immediately, because a flaw present in all three is yours.

Practice: a brief that a stranger can check

Take a real image request you have received, and rewrite it in the four layers above. Then hand the brief and three generated results to someone who was not part of the request and ask them to score each result on the four axes.

Two things usually happen. They will score at least one axis differently from you, and the reason will be a constraint you knew but did not write down. And you will find at least one axis you cannot score at all, because the brief never specified what "right" was.

Both are the exercise working. Add the missing lines to the brief, and keep the brief — it is more reusable than any individual prompt.

Common mistake

Do not assume a generated image is safe to publish because you generated it.

Four separate things need checking, and only the first is obvious. The provider's output terms, which vary and change. Any reference image you supplied, which carries its own rights regardless of how much the output differs. Whether the image contains an identifiable person or a recognisable trademark that arrived without being asked for. And whether the image makes an implicit claim — a product doing something it cannot do, or a scene that a viewer would read as a photograph of a real event.

The last one has no technical signal. It requires someone to look at the image next to the caption and ask what a reasonable viewer would conclude, which is a review step rather than a setting.