Most writers try AI on a novel, love the first three chapters, then watch it fall apart by chapter thirty: the protagonist's personality shifts, a planted clue is never recovered, two dialogue scenes read in different voices. The conclusion people draw is "AI can't write long form."
Half true. The problem isn't model quality. It's that one model is asked to carry too many jobs at once.
The context leaks
By late chapters, the first few hundred thousand words of setup — relationships, item locations, the timeline — can't all fit in one generation's context window. The model sees only the recent slice and "reasonably" invents a fact that contradicts what came before. That's not bad memory. The window physically can't hold the whole book.
The voice drifts
With no outside constraint, the model slides toward whatever tone the current chapter pulls. The sharp-tongued lead gets soft. Each drift is tiny; across fifty chapters the person "isn't that person anymore."
The foreshadowing gets lost
You plant a hook in chapter twenty, meant to pay off in chapter eighty. But when the model writes chapter fifty, nothing told it "track the open threads." The hook wasn't forgotten — it was never recorded as a thing to track.
The style cracks
A protagonist's dialogue, the narration, the fight scene often want different registers. One model writes the whole cast in one uniform voice, and it reads like one person pretending to be everyone.
The fix: separate writing from checking
Wenmai doesn't chase "one model writes a whole book." It breaks the work into checkable steps:
- Canon ledger moves setup and facts out of the model's head into a backend fact book — facts, what each character knows, numbers, items, foreshadowing status. New chapters align to it first.
- Event package designs the chapter before drafting: what event advances, what hook is planted, what old thread is closed.
- Four review roles let different models each own one lane — fact continuity, character voice, and a cold final read — instead of one model writing and editing itself.
- Mistake regression keeps a running log of past errors (the bad case, why, the right fix) and re-checks new chapters against it.
In short: a single model is fine for a scene, bad for a book. Split "write" and "check," and a million-word novel stays consistent.
The draft role (DeepSeek by default) is free and unlimited — no chapter cap. The cross-review roles and mistake regression are part of the full workflow, unlocked by the $6/mo plan. You bring your own API key; Wenmai never resells tokens.
Four steps to start
- Set the canon — two lines on the world, the lead, the core conflict.
- Design the event package — what this chapter advances, what it plants.
- Draft chapter by chapter — the free draft role writes start to end; finish first.
- Turn on review when you want it tighter — cross-review roles are optional, not a gate.
Long form is assembled, not generated. Get the draft role free, write a chapter, and you'll learn more than from ten tutorials. Start at Wenmai.