A written style specification is what makes a set of pictures look related

Individually good and collectively incoherent

Generate six illustrations for six articles over six weeks and each will be defensible on its own. Placed on one page they will disagree about almost everything: line weight, palette, how much the shapes are simplified, whether shadows exist, how much empty space surrounds the subject.

The cause is that nothing carried between the sessions except your memory of what you asked for, which is worse than you think. Style words drift, you add a qualifier one week and drop it the next, and the accumulated result is a set that reads as assembled from stock rather than commissioned.

A reader does not analyse this. They register that the site looks slightly unmaintained, in the same way a page with three fonts does, without being able to say why. Consistency is one of the few visual properties that is noticed only in its absence.

Write down the decisions, not the adjectives

The document that fixes this is short and specific. It names the medium and finish, the palette in exact terms, the line quality if there is one, the level of detail, how the subject is framed, where the light comes from, and what the background does. Six or eight statements, each of which a person could follow.

What does not belong in it is evaluative language. Modern, clean and striking describe your reaction rather than the picture, and they produce something different every week because they constrain nothing. A flat two-colour illustration on a warm off-white ground, with no gradients and a single light from the upper left, is a specification.

Include the negatives explicitly, since they are usually the decisions that get lost. No photographic realism, no drop shadows, no text within the image, no more than three colours, nothing cropped at the edge of the frame. Prohibitions are what keep a set from wandering as your prompts accumulate variations.

Fix the frame before you fix the style

Framing conventions do more for coherence than anything decorative. If every illustration shows a single subject at roughly the same scale, with similar margins and the same horizon behaviour, the set will look related even when the subjects have nothing in common.

Decide the aspect ratio once and hold it for the whole set. Mixing proportions is the fastest way to make a group look scavenged, and it also creates layout problems that get solved by cropping, which then breaks whatever compositional rule you had established.

Colour is the other structural lever. Naming a small palette and refusing anything outside it will unify a set more effectively than any amount of matching the rendering style, partly because colour is what a reader registers first and partly because it is easy to state precisely.

Reuse the block, and let the subject be the only variable

In practice this becomes a fixed block of text that begins every request, followed by one sentence describing the specific subject. That structure is deliberate: the constant part is long and unchanged, the variable part is short, and any drift is traceable to the sentence you wrote today.

Keep the block in a file rather than in your history, and note the date and tool alongside it, because outputs from the same block change when the underlying tool does. When a new picture no longer matches the set, the first question is whether the block changed or the tool did, and the file answers it.

Where a tool offers reference images or a saved style handle, use it as well as the written block rather than instead of it. The handle is convenient and opaque; the written version is what you can hand to an illustrator, adapt for a different tool, and read in a year to understand what was intended.

Accept the drift, and check against the set

Exact matching is not achievable. There will always be variation in weight and detail between images, and chasing it produces a stiff, over-specified look. The realistic target is a family resemblance strong enough that no single picture stands out.

The right check is comparative rather than absolute. Put the new image beside four existing ones at the size it will actually be used, and look for the property that is out of step — usually palette or level of detail. Judged alone, almost anything passes.

When a set has drifted too far to reconcile, regenerating the whole group from the current block is often cheaper than editing individual images towards each other, since generation is inexpensive and reconciliation is not. That is one of the genuine advantages of a written specification: it makes the whole set reproducible rather than merely documented.