People hear "use multiple AI models to write a novel" and picture four chat tabs open at once. That's not orchestration. That's four copy-paste panes.
Orchestration means each step of the work is owned by a role built for it, and the roles hand off through a shared fact book. The models are interchangeable; the system is the product.
The roles
- Canon — keeps characters, items, timeline and open foreshadowing in a fact ledger outside any model's head. Every chapter aligns to it before drafting.
- Draft — one model writes the chapter. Its only job is prose. It never edits itself.
- Cross-review — three lanes, three checks: fact and continuity, character voice, and a cold final read on a different model than the one that drafted.
- Final audit — a clean, publishable pass, not a list of suggestions.
- Mistake regression — a running log of every error caught, re-checked against every later chapter.
Why separate models at all
A single model writing and editing itself reads its own draft with the same eyes that wrote it. The drift, the contradiction, the uniform voice — all come from that loop. Separate the writer from the checker and the checker from the reader, and each can do one job well.
Why "orchestration," not "more models"
Four models in four tabs is still one person juggling. The value is the handoff: canon feeds draft, draft feeds review, review feeds audit, audit feeds regression, regression feeds the next chapter. The book gets harder to break as it grows.
What it costs
The draft role is free (DeepSeek, unlimited chapters). The cross-review roles and regression log are part of the $6/mo plan — you're paying for the method, not for model capacity. You bring your own keys (BYOK); Wenmai never resells tokens.
Start
Set a two-line canon, draft chapter one free, and turn on review when the cast gets bigger than you can track by hand. The system is there whether you use all of it or not. See how it works or the BYOK explainer.