You start a novel with a sharp-tongued sidekick and a cold antagonist with a clear motive. Forty chapters later the sidekick is soft, the antagonist forgot why they started, and two characters sound like the same person. This is the most common way an AI-written book falls apart — not the plot, the people.
The cause is structural, not a model flaw. When one model both writes and edits, it re-reads its own draft with the same habits that produced the drift. Nothing outside the model is saying "this character is not allowed to do that."
The four ways characters break
- Voice drift. The cynical lead slowly turns sincere because the recent chapters pull that way. Each step is small; across fifty chapters the person isn't that person.
- Motive contradiction. A character acts against a decision they made in chapter six, because nothing tied the later scene back to that decision.
- Knowledge leaks. Someone "remembers" a fact they were never in the room to learn.
- Uniform cast. Protagonist, villain, comic relief all get written in one house voice, so the book reads like one actor doing voices.
Fix 1: a canon ledger, not a hope
Write the non-negotiables down outside the model: who each character is, what they want, what they know, how they talk. In Wenmai this is the canon ledger. Before a chapter is drafted, the draft role reads the relevant characters from it. The model isn't trusted to "remember" — the facts are handed to it.
A ledger beats a prompt because a prompt expires when the context window fills. A ledger is queried per chapter, so chapter eighty still sees chapter three's setup.
Fix 2: a separate voice-checking role
Don't let the writer grade its own voice. Wenmai gives the character-voice role to a different model pass: it reads the new chapter and checks whether each character sounds like themselves, quotes the line that broke, and proposes the fix. The draft model never sees that critique while writing, so it can't over-correct toward "safe."
Fix 3: a mistake-regression log
The first time a character contradicts an earlier fact, log it: the bad line, why it happened, the right version. Wenmai re-checks every later chapter against that log. The book gets harder to break the longer it gets, instead of easier.
What this looks like in practice
| One model, one chat | Wenmai | |
|---|---|---|
| Character facts | In the model's memory | In a canon ledger, queried per chapter |
| Voice check | You, by hand | A separate role, different model |
| Contradictions | Caught by you, late | Logged and re-checked every chapter |
| 50-chapter book | Usually drifts | Stays consistent |
Start small
You don't need the full system on chapter one. Set the canon with a few lines per main character, draft freely on the free role (DeepSeek, unlimited chapters, no card), and switch on voice-checking when the cast grows past what you can track by hand. The $6/mo plan unlocks the cross-review roles and the regression log.
Consistency isn't a talent you hope the model has. It's a structure you build. See how Wenmai works or read why one model can't finish a novel.