Style Transfer: Teaching AI Your Voice

# Style Transfer: Teaching AI Your Voice

By the time a writer has spent a few weeks with a large language model, a familiar complaint begins to surface. The prose is *fine*—grammatical, fluent, occasionally elegant—but it does not sound like anyone in particular. Read three AI-assisted paragraphs and you can already hear the telltale cadence: the balanced clauses, the polite transitions, the careful hedges, the abstract nouns where a concrete one would bite harder. It is the voice of a very well-trained committee, and the committee is not you.

In the last two essays we have treated prompting as composition: ["Prompt Engineering Is Not Coding — It's Composition"](/harrys-desk/prompt-engineering-is-composition) argued that a prompt is a rhetorical brief, not a function call, and ["Chain-of-Thought, Few-Shot, and Structured Prompting"](/harrys-desk/chain-of-thought-few-shot-structured) showed how examples, reasoning steps, and output formats can shape what the model produces. This week we take the next logical step. If prompting is composition, then *style transfer* is the art of making the model compose in a voice that is recognizably your own.

This is not vanity. Voice is the writer's signature, the thing that turns information into experience and argument into relationship. When ["The Writer's Dilemma: Why AI Changes Everything"](/harrys-desk/the-writers-dilemma-why-ai-changes-everything) argued that the writer's job shifts from production to curation, it left unanswered a deeper question: curation of *what*, exactly, if every paragraph sounds interchangeable? The answer is that the writer must become the curator of voice. The machine supplies velocity; the writer supplies the sound that makes a reader trust, remember, and feel.

What Voice Actually Is

Writers talk about voice as if it were a mystical property, but it is made of observable parts. Diction, syntax length, rhythm, figurative density, rhetorical stance, preferred transitions, tolerance for fragments, appetite for abstraction—these are all patterns that can be described, sampled, and imitated. A large language model is, at its core, a statistical pattern-matching engine. That means it is extraordinarily good at imitation once you point it toward the right patterns.

The catch is that most writers do not know their own patterns well enough to describe them. We write by ear. We know when a sentence sounds like us and when it does not, but we often struggle to say why. Teaching a model your voice therefore becomes a strangely useful exercise in self-study. You are forced to reverse-engineer your own prose, which is one of the fastest ways to become a more deliberate writer.

Start by gathering a corpus. Five hundred to a thousand words of your own prose is usually enough to begin, though two thousand is better. Choose pieces that feel *most like you*—not your most polished public work, necessarily, but the writing where your instincts ran ahead of your editing. Essays, journal entries, letters, unpublished scenes, even well-written emails can work. What matters is that the sample captures the unguarded rhythm of your mind.

Once you have the corpus, read it as a stranger would. Underline recurring tics. Do you begin sentences with conjunctions? Do you prefer short Anglo-Saxon nouns or Latinate abstractions? Do you stack clauses or cut them off? Do you use metaphor sparingly or lavishly? Do you address the reader directly, or observe from a distance? Do you qualify your claims with *often*, *usually*, *in my experience*, or do you state them flatly? Every answer is a lever you can hand to the model.

The Frame: Naming the Voice Before You Ask for It

Style transfer begins with a frame. In our earlier vocabulary, a *frame* tells the model what kind of text it is being asked to produce: a legal memo, a lyric essay, a technical manual. For voice transfer, the frame must be more specific. You are not asking the model to write "an article." You are asking it to write *like someone*.

A weak frame sounds like this: "Write in my style." A strong frame sounds like this: "Write in the voice of a writer who favors short sentences, concrete nouns, occasional fragments, direct address to the reader, and a skeptical but warm rhetorical stance. Avoid abstract hedges. Prefer active verbs. Use metaphor only when it clarifies something that literal language cannot."

Notice the difference. The first frame offers a label with no content. The second offers a profile. Large language models do not have a private mental image of "your style." They have only the text you give them. The more granular your description, the closer they can come.

Even better is to combine the profile with exemplars. This is where few-shot prompting becomes style transfer. Take two or three short passages from your corpus and include them in the prompt with a label: "Here are examples of my prose." Then ask the model to produce new text that matches the same voice while treating a different subject. The model will attend to surface patterns—sentence length, diction, transitions—but also to deeper structures: how you move from observation to judgment, how you handle uncertainty, how you pace revelation.

The Constraint: Teaching the Model What to Avoid

Imitation is only half the task. The other half is avoidance. Every writer has not only a voice but an anti-voice: the phrases they would never use, the rhythms that feel foreign, the tonal mistakes that make them wince. A model trained on the whole internet will default to the middle of the distribution unless you pull it elsewhere. That middle is exactly what you are trying to escape.

The most effective way to teach avoidance is to list the patterns you do *not* want. This feels pedantic, but it works. "Do not use the following phrases or their close variants: *delve into*, *in the realm of*, *it is important to note that*, *as a writer*, *the ever-evolving landscape of*." "Do not begin consecutive sentences with the same construction." "Do not resolve every paragraph with a summarizing aphorism." "Do not use intensifiers like *very*, *truly*, or *deeply* except in dialogue."

These instructions are not about being a scold. They are about narrowing the model's probability distribution. Every phrase you forbid removes a likely default and forces the model to reach further down the stack for an alternative. The alternative may still be imperfect, but it will be less generic.

Another useful constraint is length. Most AI-default prose expands to fill a predictable shape: an opening hook, three balanced points, a closing reflection. If your natural voice is more compressed or more sprawling, say so explicitly. "Keep paragraphs to three or four sentences. Let some paragraphs be a single sentence." Or: "Write long, winding paragraphs that build momentum through accretion rather than bullet-point clarity." The model can honor either instruction; it simply needs to know which one you inhabit.

The Dialogue: Revision as Voice Calibration

Voice transfer rarely succeeds on the first pass. The first output may capture your sentence length but miss your irony. The second may get the irony but add a phrase you would never use. This is not failure; it is the normal texture of co-composition.

The best method is to treat the model as a student who is trying to learn your voice and needs specific correction. Quote offending sentences and say what is wrong with them. Offer a rewrite in your own hand and ask the model to diagnose the difference. Over three or four rounds, the model will converge toward a much closer approximation.

Here is a technique I use regularly. After the first draft, ask: "Which sentences in this paragraph sound least like the exemplars I gave you? Explain why, then rewrite them." The model is surprisingly good at identifying its own deviations when given a clear comparison point. The explanation it produces is also useful to you: it surfaces patterns in your own writing you may not have noticed.

Do not be afraid to overwrite sentences yourself. Voice transfer is not a magic trick in which the machine reproduces you perfectly. It is a collaborative sculpting process. The model proposes; you refine. The closer the proposal, the less refining you have to do, but the final decisions must still be yours. If you accept every AI sentence without revision, you are not transferring your voice; you are renting the model's.

Voice vs. Persona: Do Not Confuse the Two

A warning is necessary here. Voice is not persona. A persona is a mask: the cynical detective, the enthusiastic explainer, the stern academic. A voice is the recognizability that persists beneath any mask. You can write a horror story, a business memo, and a love letter in the same voice if your diction, rhythm, and stance remain consistent.

Models are excellent at persona and only moderately good at voice. Persona is easier because it maps cleanly onto genre conventions: the detective uses short sentences and hardboiled metaphors; the explainer uses analogies and enthusiastic transitions. Voice is harder because it is idiosyncratic, inconsistent, and emotionally specific.

This is why generic prompts like "write like Hemingway" work better than "write like me." Hemingway is a persona the model has encountered thousands of times. You are a voice it has never heard. The only way to teach it is to supply the evidence and the instructions in the same prompt.

When I ask a model to adopt a persona, I am usually producing a sketch or a parody. When I ask it to adopt my voice, I am doing something more intimate. I am asking the machine to sound like the part of me that chooses words when no one is watching. That part is worth protecting.

Style Transfer in Practice: A Worked Example

Let me show you how this looks in practice. Suppose I want the model to write a paragraph about revision in my own voice. I might prompt it as follows:

> Write a paragraph about the pleasure of revision. Use the voice of a writer who favors short sentences, concrete nouns, direct address, and occasional wit. Avoid abstract nouns like *process*, *journey*, and *craft* unless no plain alternative exists. Do not use phrases like *delve into* or *it is important to note*. Here are two examples of my prose for reference: [paste two short passages]. Now write the paragraph.

The first output might be acceptable but slightly off. Perhaps it writes: *"Revision is where the real magic happens, because this is the moment when you strip away everything that isn't your voice and find the sentence that was waiting underneath."*

I would respond: *"The phrase 'the real magic happens' is not mine. I do not use the word 'magic' metaphorically. Also, 'this is the moment when' is throat-clearing. Try again, more compressed."*

A better output might be: *"Revision is subtraction. You cut the sentence that sounded clever at midnight and find the shorter one underneath. The reader never knows what you removed. That is the pleasure."*

Closer. Still not exactly me, but now I can edit rather than rewrite. The model has done the heavy lifting of compression and concrete diction; I supply the final calibration.

The Ethics of Synthetic Voice

Style transfer raises a question we will return to throughout this series: when the model can imitate your voice, who owns the result? If I feed the model my prose, refine its output, and publish it under my name, the work is mine in every meaningful sense. I chose the subject, set the constraints, corrected the deviations, and made the final decisions. The model was a tool, albeit a sophisticated one.

But the line grows slippery if I begin to use the model to imitate *someone else's* voice. The internet is already full of synthetic Hemingways and fake Shakespeares. That is usually harmless parody or pedagogy. It becomes harmful when the imitation is used to deceive, to borrow authority, or to pass off generated text as another person's authentic statement. ["The Symbiotic Model — Human + AI, Not Human vs AI"](/harrys-desk/the-symbiotic-model) argued that the moral unit of writing is the human intention behind it. Voice transfer puts that principle to the test. Imitate your own voice freely. Imitate others' voices only with consent, disclosure, or clear artistic purpose.

There is also a subtler danger: the model's imitation can become a cage. If you always prompt in your own voice, you never experiment. One of the great gifts of AI is the ability to try on voices you would never naturally use—the terse journalist, the ornate essayist, the clinical analyst. Those experiments can feed your own prose with new rhythms and new ideas. Do not let style transfer become a fortress. Let it be a home you can leave and return to.

Building a Voice Library

Over time, you will want to formalize what you have learned. I keep a simple voice library in my notes: a single document that lists the elements of my preferred style, the phrases I want to avoid, and three or four exemplar passages for different registers. When I start a new project, I paste this document into the model as context. It is faster than reconstructing the profile from scratch every time.

A voice library typically includes:

• •A one-paragraph description of the voice, written in the voice itself.
• •A list of five to ten stylistic signatures: sentence length, diction level, figurative habits, transitions, stance toward the reader.
• •A forbidden-phrase list, updated whenever the model produces something that makes me wince.
• •Three exemplars at different lengths: a short paragraph, a medium passage, and a longer passage with variation.
• •A revision instruction: how I want the model to respond when I say a sentence is off.

This library is a living document. Every article I publish teaches me something new about my own tendencies. The model becomes, paradoxically, a mirror that helps me know my own sound better.

Why This Matters for the Series

We are still in the foundations section of this masterclass, and there is a reason style transfer appears early. Everything that follows—fiction, non-fiction, journalism, technical writing, biography, business prose—depends on the writer being able to maintain a consistent and appropriate voice. A novelist who cannot teach a model her narrator's cadence will spend hours line-editing machine-generated dialogue. A journalist who cannot enforce a clipped, factual style will watch AI padding seep into his reporting. A technical writer who cannot constrain the model's impulse toward elegance will produce documentation that is pleasant to read and useless in a crisis.

Voice is not ornament. It is the contract between writer and reader. It tells the reader what to expect, how to listen, and how much to trust. If you hand that contract to a model without instruction, the model will sign it in the blandest handwriting available. Your job is to teach it your signature.

For Next Time

On Wednesday we move from voice to *genre calibration*: the art of tuning a model for fiction versus non-fiction, essay versus report, lyric versus argumentative prose. We will look at how the same facts can be rendered in a dozen different genre shapes, and how to prompt the model into the right one without losing your own sound.

---

*Harry Mercury, Editor in Chief* *The SMF Works Project* *Week 4, Article 1*