Skip to main content
A variant is one dimension of customisation. Language, brand, audience, and persona are the usual ones. Each dimension has options, and each option carries fields describing it:

​
One option per dimension, per session

When a session starts, exactly one option is chosen per dimension — and each combination produces its own separate output. Two languages and two audiences is four generations per record, and four times the cost.

​
Dimension or field?

The deciding test: does this thing need its own output candidate per option?If yes, it is a dimension. If it is additional information that shapes the same output — tone notes, reading level, terminology — it is a field on an existing option.
Getting this wrong in the direction of too many dimensions is expensive, because the combinations multiply. Keep to three or four dimensions at most.

​
Why variants, and not the prompt

Variant fields are injected into the generation prompt and into every AI quality check’s context. That is what makes them different from special instructions or prompt text. Anything that must be quality-checkable — brand voice, terminology, audience traits — belongs in a variant, because it is the only way the checker knows what the writer was told.

​
Curating fields

  • Two or three fields per option is the safe baseline
  • Four to six needs testing
  • Seven or more risks the model quietly dropping some
Every field is a constraint the model has to hold simultaneously. If removing a field would not change the output, remove it.
Keep the default instruction unless you have a specific reason to change it.

​
Scoping rules to variants

A quality rule can be scoped to specific options — a forbidden-terms list that applies only to Danish, a formality check only for German. A rule that does not match the session’s combination is not even shown to the generator. Quality checks