Building a style profile

What to feed it

Paste in pages you are happy with — your own, from this project or an older one. Two to five thousand words is enough to get a usable profile; more helps.

Feed it prose, not notes. The extractor is looking at rhythm and diction, and an outline has neither.

What comes out

Generation produces two things:

1. A guide — a few paragraphs describing the voice in terms the model can act on. You can edit it. It is prose, not a config file.
2. Assets — individual extracted items: characteristic phrasings, imagery, dialogue habits, sentence patterns. Each carries tags (style, scene, function) and a caution level.

Only one profile is active per project at a time.

How assets reach the model

Not all of them, and not always. At generation time the assets are selected by the scene's tags — a tense confrontation pulls different phrasings than a quiet interior scene — and the selection policy decides how many and how strongly.

You can preview exactly which assets would be included before generating. That preview costs nothing; it is pure assembly.

Assets carry a use count, so an asset that keeps showing up in every scene is visible as such — that is usually the sign to disable it, not to keep it.

When it goes wrong

If generated prose starts sounding like a parody of itself, the usual cause is too many high-caution assets active at once. Turn the selection policy down before you start deleting assets.