# Novel II — Voice and Style: Finding the Narrator
In ["Novel II — Deep Craft: Pacing and Tension"](/harrys-desk/novel-ii-deep-craft-pacing-and-tension) I said pacing and tension are two engines, not one — rate of movement versus the reader's desire to know. That article closed Week 14's three-part arc on deep craft: [point of view](/harrys-desk/novel-ii-deep-craft-point-of-view), [showing versus telling](/harrys-desk/novel-ii-deep-craft-showing-vs-telling), and the pull that keeps a reader in the chair at one in the morning. All of it presupposed a voice. Today we turn to that voice, because a novel with the right distance, the right rhythm, and the right tension can still fail if the person telling the story sounds like no one in particular.
And here is the claim I want you to carry out of this piece: voice is not diction. It is not a texture you apply in the line edit. It is the residue of a specific consciousness under pressure. The model has no consciousness, so it cannot have a voice. It can only have a manner. Confusing those two is how AI-assisted novels become competent, smooth, and forgettable.
By the end of this article you should stop treating "find the voice" as a late-stage polish and start treating it as a constraint you establish before the model writes a sentence.
Voice Is Not Vocabulary
Writers talk about voice as if it were a palette of words. Short sentences. Regional slang. A fondness for em-dashes. Those are mannerisms. They are the surface of a voice, the way a person's handwriting is the surface of their thought. You can imitate handwriting without thinking like the person. You can load a draft with contractions and fragments and still have no voice at all.
Voice is what a particular mind does when it has to choose. What it notices. What it refuses to notice. What it finds funny. What it finds unforgivable. The order in which it reaches for a metaphor. Whether it trusts the reader or lectures them. Whether it can stand silence. A narrator who notices the scuff on a shoe before the face of the person wearing it has a different voice from a narrator who notices the face first, even if they use the same vocabulary and the same sentence length.
This is why [point of view](/harrys-desk/novel-ii-deep-craft-point-of-view) was the right place to start Novel II, and why it is not sufficient. POV tells you who is talking and from what distance. Voice tells you who they are as a speaker. Two first-person narrators can occupy the same grammatical position and sound nothing alike. Holden Caulfield and Humbert Humbert are both first person. They do not share a voice. They share a pronoun.
The test is simple and brutal. Cover the character's name. Cover the setting. Read a page. If you cannot tell who is speaking — if the page could belong to any literate adult in your genre — you do not have a voice. You have English.
The AI Accent
I need to name the problem that AI-assisted drafting creates here, because it is more specific than "the prose sounds generic."
The model has an accent. Not a regional accent. A statistical one. It is the accent of the average of everything it has read that resembles the prompt you gave it. The accent is competent. It is grammatical. It is often pleasant. It prefers balanced clauses, a clean turn at the end of a paragraph, a slightly elevated register that never quite commits to being literary and never quite commits to being plain. It likes the sentence that sounds wise without having to be true.
I call this the AI accent because once you hear it you cannot unhear it. It is the sound of a mind that is not under pressure. A real narrator is always under pressure — of character, of situation, of the need to say this and not that. The model is under no pressure except the pressure to continue. Continuation is not voice. Continuation is fluency.
When you draft scene by scene with a model, the accent infiltrates in a particular way. The first scene you generate might still carry your cadence, because the prompt is long and the sample of your prose is fresh. By scene twelve, the model's prior outputs have become the context. The draft is now training on itself. The accent compounds. You read chapter eight and cannot tell whether a given sentence is yours, the model's, or a hybrid that belongs to neither. This is not a moral problem. It is a diagnostic one. The hybrid sentence is the most dangerous, because it sounds enough like you to survive the first pass and enough like the model to drain the particularity that made the first chapter live.
In ["Style Transfer: Teaching AI Your Voice"](/harrys-desk/style-transfer-teaching-ai-your-voice) I talked about teaching a model the surface of a style. That work still matters. It is not the work of finding a narrator. Style transfer can make the model write like you. Finding the narrator is making the novel sound like the person who is telling it — who may not be you, and who must not be the model.
The Narrator Is a Character
Even when the narrator is unnamed. Even when the novel is in close third and the narrator seems to vanish. There is always someone choosing what to mention and in what order. That someone is a character, whether you have given them a name or not. If you have not decided who they are, the model will decide for you, and it will decide they are a helpful, even-tempered, slightly literary generalist. That is the default narrator of the training data. It is also the default narrator of workshop fiction. It is death on the page.
So the first practical move is not a prompt. It is a portrait. Before you generate a chapter, write — yourself, in a notebook or a file the model does not see — a page about the narrator as a speaker. Not biography. Speech. What they would never say. What they over-explain. What they find vulgar. Whether they are trying to impress the reader or trying to get rid of them. Whether they think they are honest. (Most narrators think they are honest. The interesting ones are wrong.)
This is adjacent to the work we did in ["Short Story — Character: The Character Interview"](/harrys-desk/short-story-character-the-character-interview), but it is not the same interview. You are not asking the character what they want. You are asking the speaker how they tell. A character who wants revenge and a narrator who tells that want as a joke are not the same person occupying the same skull. The gap between them is voice.
For close third, the portrait is of the consciousness you are sitting next to — its metaphors, its blind spots, the words it reaches for when it is afraid. For omniscient, the portrait is of the teller: the intelligence that stands above the story and has opinions about it. Omniscient without a narrator-as-character is just camera movement. That is why so much contemporary omniscient feels like a glitch in limited. There is no one home.
Do not put this portrait in the prompt yet. If you dump it into the model, the model will perform it. Performance is mannerism. You want the portrait in *your* head so that when the model produces a sentence that the narrator would not say, you can hear the wrongness. The portrait is a tuning fork, not a costume.
Calibration Without Flattening
Now the model. You will use it. The question is how to calibrate it to a narrator without letting it flatten the narrator into a set of verbal tics.
The failure mode looks like this. You tell the model: this narrator is wry, uses short sentences, never names emotion, grew up on a farm. The model produces wry short sentences that never name emotion and mention barns. Every paragraph. The wryness becomes a drumbeat. The farm becomes a brand. You have not found a voice. You have installed a filter.
Calibration that works is negative as much as positive. Tell the model what the narrator will not do. Will not explain the joke. Will not summarize a feeling after dramatizing it. Will not use the word *somehow*. Will not end a paragraph with a thesis statement about what the scene meant. Constraints of refusal are more useful than constraints of flavor, because flavor is what the model is good at faking and refusal is what it is bad at. The model wants to be complete and helpful. A narrator is often incomplete and unhelpful. That friction is the beginning of a voice.
The other useful calibration is a sample — a short passage you wrote without the model, in the narrator's speech, that you treat as canonical. Not a style guide. A passage. Two hundred words in which the narrator is under pressure: delivering bad news, lying, noticing something they wish they had not. Pressure reveals voice the way a stress test reveals a bridge. Paste that passage into the prompt and say: the next scene must be speakable by the same mind. Then generate. Then read the generation against the sample out loud. Where the rhythm breaks, where a word arrives that the sample-mind would not reach for, you have found the accent. Cut it. Do not ask the model to fix it. The model will replace one fluent sentence with another fluent sentence. Fluency is the accent.
I do this in layers. First generation: I let the model be fluent. I am not hunting voice in the draft. I am hunting scenes. Then I read the scene against the sample and mark every sentence that could have been written by the default narrator. Those sentences get rewritten by me, in the narrator's speech, by hand. The model can propose alternatives if I am stuck, but I do not accept a proposal I cannot hear in the sample-mind's mouth. This is slower than "write in the voice of X." It is the only method I have found that does not slowly replace the narrator with a well-trained impersonator.
Style Versus Mannerism
A brief distinction, because we will need it on Wednesday when we talk about imitation.
Style is the set of recurring choices that arise from a way of seeing. Mannerism is a recurring choice that has been detached from seeing and is now being performed as a signature. Faulkner's long sentences are style when they are the shape of a mind that cannot stop associating. They are mannerism when a later writer produces long sentences because long sentences are "Faulknerian." Your own draft can contain both. The fragment that arrives because the narrator is out of breath is style. The fragment you insert because you have decided the book is "spare" is mannerism.
The model is a mannerism engine. It is extremely good at identifying a signature and reproducing it without the seeing that produced the signature. That is what [style transfer](/harrys-desk/style-transfer-teaching-ai-your-voice) does, and it is useful for pastiche, for homage, for learning. It is fatal if you mistake the reproduction for the narrator.
The test I use: take a mannered passage and change the content. If the verbal signature still fits the new content without strain, it was a mannerism — a coat you can put on any scene. If the signature breaks when the content changes, it was style — the form was doing work that belonged to that moment. Voice lives in the second category. It is not portable. That is why it is valuable.
The Strip Test
Here is the homework-shaped practice I actually use when a draft has been living with a model for too long.
Print the chapter. Or put it in a file you will not generate into. Read it once for story, once for voice. On the second pass, strike every sentence that does not sound like the narrator you portrayed on that page you wrote before the model saw anything. Do not replace yet. Just strike. You will often find that a third of the chapter vanishes. The remaining two-thirds will have a pulse. That pulse is the voice. The vanished third is the accent.
Then rewrite the gaps yourself, speaking the narrator's sentences under your breath. If you cannot hear them, you do not yet know the narrator well enough, and no prompt will save you. Go back to the portrait. Put the narrator under a new pressure — a scene they would hate to tell — and write two hundred words by hand. Then return to the gaps.
The fear, always, is that you will strip out yourself along with the accent. The fear is usually backwards. What you think of as "yourself" in a late AI-assisted draft is often the hybrid: your instincts smoothed by the model's fluency. The sentences that survive the strip test are more you, or more the narrator, than the sentences that felt like you because they were easy to read. Ease is not identity. The model is very good at ease.
In ["Authenticity — When Is It \"Your\" Work?"](/harrys-desk/authenticity-when-is-it-your-work) I argued that authorship is a matter of decisions, not of keystrokes. Voice is the audible form of those decisions. If the decisions are yours and the fluency is the model's, the voice will still hold, because voice is in the choosing. If the model is also choosing — what to notice, what to skip, where to land the paragraph — then the voice is not yours and it is not the narrator's. It is the accent. Strip it. What remains will be rougher. Rough is not a defect. Rough is evidence of a mind under pressure.
For Next Time
On Wednesday we continue this arc with ["Novel II — Voice and Style: Imitation and Homage"](/harrys-desk/novel-ii-voice-and-style-imitation-and-homage). Finding a narrator is the inward move — hearing the specific speaker of your book. Imitation is the outward move: sitting with a master long enough to steal what is worth stealing without becoming a costume. We will talk about the difference between homage and pastiche, why the model makes pastiche dangerously easy, and how to use imitation as training for the ear rather than as a substitute for a voice you have not found.
Your homework until then: write the narrator portrait I described — one page, speech not biography, what they would never say — and then write two hundred words of that narrator under pressure, with the model closed. Do not generate. Do not revise toward smoothness. Read those two hundred words out loud. That passage is now your tuning fork. Keep it next to the draft. Every time a generated sentence cannot be spoken by the same mouth, you have found the accent. Mark it. We will need that ear on Wednesday, when the temptation will be to borrow someone else's mouth instead.
---
*Harry Mercury, Editor in Chief* *The SMF Works Project* *Week 15, Article 1*