Keeping Your Voice Consistent When Writing With AI: A Voice Profile That Actually Works

The fix has two parts. First, give the model your voice in a form it can actually use, which is not "write in a literary style" but a profile with samples, rules and a banned list. Second, check for drift on a schedule, because even a good profile is only as strong as the writer's editing. This article covers both.

What "voice" is, operationally

For the purpose of instructing a model, voice decomposes into things you can name:

A model given these rules will follow most of them most of the time. A model given "write in an evocative literary voice" will produce its default, which is the same voice it gives everyone.

The voice profile

Here is the profile format I use. It goes into the style section of the story bible, and it is attached to every drafting generation.

VOICE PROFILE — [Novel title]

Fixed: close third person, past tense, one POV per scene.

Rhythm: sentences mostly 8–15 words, varied; fragments allowed after a long
sentence; paragraphs uneven, 1–5 sentences; a one-line paragraph marks a turn.

Distance: narrator knows only what the POV character knows and notices only
what she would notice. No omniscient asides.

Diction: plain, concrete. Anglo-Saxon over Latinate where both exist. Trade
vocabulary (archive, seal, vellum, ledger) used exactly.

Figures: at most one metaphor or simile per page, drawn from paper, ink,
weather or the body. Never from technology or food.

Emotion: never named. Rendered as what the body does or what the character
looks at instead.

Thought: free indirect. No italics, no "she thought."

Dialogue: "said" or no tag. Beats carry attribution. No adverbs on tags.

Punctuation: no em dashes. Commas and full stops. Semicolons rare.

BANNED: a testament to; couldn't help but; a mix of X and Y; in that moment;
little did she know; felt a surge/wave/pang of; the weight of; delve; tapestry;
testament; palpable; a sense of; found herself; realised that; eyes widened.

SAMPLES (verbatim, do not quote back):

[Action sample, ~120 words]

[Interior sample, ~120 words]

[Dialogue sample, ~120 words]

Three things about this format that matter.

The samples are verbatim and chosen for range. Three passages from your own work, each showing a different mode. The model imitates samples more faithfully than it follows rules, so the samples are the load-bearing part. If you do not yet have prose you are happy with, write the three samples first, slowly, by hand. They are the most important 360 words of the project.

The banned list is specific. Models have defaults: constructions that appear in their output far more than in published fiction. Listing them explicitly reduces their frequency substantially. Your list will grow as you notice your model's particular habits. The AI prose editing guide has a longer starting list.

The rules are stated as constraints, not aspirations. "No em dashes" is followable. "Elegant punctuation" is not.

Per-character voice

Narrative voice is one problem; each character's speech is another. A model given no guidance writes every character with the same rhythm and the same willingness to explain themselves. The fix is a speech section in each character card:

TEO — speech
Long sentences that trail off; starts with "Look" when about to lie; uses
Mara's full name when angry and nothing at all when hurt; never asks a direct
question; jokes to end conversations.

Sample lines:
"Look, it isn't that I didn't want you to know. It's that knowing would have
meant—well. You'd have come back sooner, and I wasn't ready to be looked at."
"Theodora Vell, you will sit down."

Two or three sample lines per character do more than a paragraph of description. When drafting a scene, the cards for the speaking characters are in the context alongside the voice profile, and the dialogue guide covers how to make the differences audible.

How the tools handle it

As of September 2026, the major tools attach style in different ways.

• Sudowrite has a Style field in the Story Bible and "Prose Modes" (including its in-house Muse model) that shape output. The Style field is where the profile goes; it is attached to generations automatically.
• Novelcrafter uses prompt personas, presets and fully editable prompts, so the profile can be placed wherever you like in the template. Powerful; requires setup.
• NovelAI shapes voice through modules, presets and the Author's Note, which are closer to sampling settings than to a profile. Samples in the Memory field help.
• General chatbots need the profile pasted every session or placed in a custom instruction, which has a length limit that a full profile may exceed.
• SparkLore has a style entry in the bible with the sections above, attached to every drafting action, and per-character speech notes on character entries that are included when that character is in the scene.

Whichever you use, the test is whether the profile is present in every generation. If it is present only at the start of a session, drift begins with the second request.

The drift checks

A profile reduces drift; it does not remove it. Two checks, on a schedule.

The read-aloud check (weekly)

Take a paragraph you are proud of from chapter one and one from the current chapter. Read both aloud. If they sound like two writers, compare:

1. Sentence length: count the words in ten consecutive sentences from each.
2. Paragraph length: count sentences in ten consecutive paragraphs.
3. Distance: does the later passage explain what the character feels?
4. Vocabulary: any words in the later passage the earlier writer would not have used?
5. Figures: more, fewer, from a different domain?

Usually one of the five has moved. Fix that one across the recent chapters and add a line to the profile that names the failure.

The model check (per chapter)

Give the model both paragraphs and ask:

Here is a paragraph from chapter 1 and one from chapter [n]. Do they read as
the same writer? List specific differences in sentence length, rhythm, distance
from the POV character, vocabulary and figurative language. Do not rewrite.

The model is a competent comparer, and it will name differences you have stopped seeing. It will also occasionally invent differences; use it as a prompt for your own reading, not as a verdict.

The rewrite discipline

The profile is only half of the method. The other half is that AI output is a scaffold, and the voice is applied in the rewrite. A concrete rule I keep: every sentence in a generated passage is either rewritten or deleted before the next generation. Not "lightly edited." Rewritten, so that the rhythm and word choice pass through my hands.

This sounds slow. It is faster than drafting from nothing because the scaffold has already solved blocking, order and beat, and it is much faster than trying to fix a generic voice in revision, when it has spread across sixty thousand words.

A before-and-after, with the profile above in the context:

Generated:

Mara felt a wave of unease wash over her as she stepped into the archive. The familiar smell of old paper and dust was a testament to the years she had spent here, and in that moment she couldn't help but wonder if she had made a mistake.

Even with the profile, the banned constructions leaked through (this happens; the list reduces, it does not eliminate). Rewritten:

The archive smelled the way it always had, of paper kept dry at some cost. Mara stopped inside the door. Ten years of her life were shelved in this room under other people's names, and she had come back to add one more.

Nothing in the rewrite is in the original except the archive and the smell. The scaffold told me the scene opened with her entering and feeling doubt. The rewrite decided what the doubt was made of.

When the voice is not yet yours

A frequent situation: a writer early in their first novel does not yet have a settled voice to profile. Two suggestions.

Write the three samples by hand, slowly, over a few days, before any AI drafting. Treat them as the standard. It is easier to find a voice in 360 words than in 80,000.

And do not let the model choose for you. If you draft the opening with AI and adopt whatever voice it produced, you will be writing in the model's default for the rest of the book, and so will everyone else who did the same thing. A voice is a set of choices; make them before the tool makes them for you.

Try it in SparkLore

SparkLore's bible has a style entry with the sections in the profile above, including a banned list and three verbatim samples, and it is attached to every drafting action along with the speech notes of any character in the scene. The read-aloud check has a one-click version: select a paragraph from any two chapters and ask for the voice comparison. To test how much a profile changes output before creating a project, the free AI story continuation tool accepts a style sample alongside the passage; run the same passage with and without it. Sign up to keep the profile attached automatically.