Voice is what you won't say as much as what you will

Voice is defined as much by what a writer refuses as by what they choose. That's the claim this essay makes. Build a voice guide only from finished work and it begins with the yes-pile: the pieces published, the drafts kept, the choices made and left standing. The no-pile can stay invisible unless you record it. And the no-pile is where voice lives as much as anywhere else.

• A style guide captures what you write about, how you structure it, the phrases you reach for.
• Unless you add it deliberately, a guide built from finished work may miss the phrase you'd never reach for.
• The gap between those two things is small until it isn't.

What everyone is trying to build

The style guide approach to AI-assisted writing is the obvious move, and the people building them aren't wrong to try. You document what you write about. You note your sentence length preferences, your structural habits, the recurring phrases a reader would recognize. You paste in pieces you're proud of, tag the patterns, tell the tool: this is how I write, stay close to this.

The logic holds. If the problem is that AI drafts don't sound like you, the solution is to give the tool more information about what you sound like. More context. More examples. A richer, more precise description of the voice. The style guide is that description made explicit.

Writers who do this carefully can get drafts that feel closer. They build detailed voice documents and treat the prompt as a craft problem worth solving. The tool may reach less often for its defaults. It may pick up some of the rhythm and stay nearer the territory.

That's a useful place to start: before you use AI to help you write, give it examples of how you write.

Why a style guide can help

What a style guide actually captures is more than most writers expect. Topics, yes. But also structure: the way you tend to open a piece, where you put the turn, how you close. Sentence-level texture: whether you run long or short, whether you use fragments, whether you like a list. Register: formal or plain, warm or dry, close or at a remove.

A reader who knows your work would recognize those things. A tool given examples may produce a more familiar shape or pick up phrases you favor. You still have to read what came back.

That's the win a style guide can deliver. It moves the tool from generic toward adjacent. For some writing tasks, adjacent is close enough to make the next revision useful.

The problem only surfaces at the edge, at the specific moment when adjacent isn't enough.

The phrase that doesn't belong

Joshua Seiden documented a useful experiment: he gave Claude raw notes and a style guide, then asked readers to compare the resulting post with his own writing. His account and the linked draft are here. The experiment makes the comparison visible; it doesn't establish that a guide can never record what a writer rejects.

Consider a teaching example, not a quotation from Seiden's draft: a quiet, reflective piece ends with “Smash the heart if you're figuring it out, too!” The structure might be plausible. The invitation might still be a phrase the writer would never use.

The question is not just whether the guide captured familiar moves. It's whether it told the tool which plausible moves don't belong.

What the yes-pile leaves out

Here's what a guide built only from finished work can leave out.

Build a guide only from a writer's finished work and you begin with the things the writer chose to write and chose to keep. The published pieces. The drafts that made it through. The sentences that survived revision. That's the yes-pile: everything the writer said yes to, accumulated.

Those examples give a drafting tool a pattern to follow. That isn't the same as giving it a record of every edit you rejected.

What those examples may not show is the no-pile. Every phrase rejected before the draft was finished. Every register abandoned mid-sentence. Every instinct that said not that before a reason could be named. The piece that got written to the end and deleted. The edit that removed a sentence that almost made it. Those are refusals, and they don't appear in the yes-pile because that's what it means to refuse them.

The no-pile can be substantial. You don't need to calculate its size to see the limitation of working from finished output alone.

Voice is shaped by both piles. If you supply only one, you leave the other for the revision.

Voice is what you won't say as much as what you will

Writing homogenization is the tendency of AI-assisted prose to flatten toward the same register, the same phrasing, the same moves. It is usually described as a problem with insufficient personalization. The fix is more context, better prompts, a richer style guide.

That is one useful diagnosis. Another question is whether the guide records your refusals. A sentence can be plausible in your general register and still be a move you would reject. More examples of what you kept may not explain why you would reject it.

Voice isn't only the accumulation of distinctive choices. It's equally the accumulation of consistent refusals. The register you've never used. The phrase you've edited out every time it appeared. The move that feels wrong before you can name why. Those refusals can be invisible in the finished output, but they shape every sentence that does appear. They're the negative space the yes-pile is cut against.

Think of five things you would never write. Not topics. Moves. Registers. Phrases. Punctuation habits. A sentence structure you've avoided for years without quite deciding to.

Write the list privately. Let it inform what you keep or change.

Notice how specific they are. That specificity is what voice is made of as much as anything you've chosen to do. The authorship contract, the implicit promise you make to readers when your name goes on a piece, is kept partly through the choices they can see and partly through the refusals they never notice because the sentences they'd have noticed aren't there.

What changes if you accept this

The most useful thing this frame changes is how you read an AI-assisted draft.

The standard read is: does this sound like me? You're checking for presence: your phrases, your rhythm, your structural moves. When you find them, the draft feels close. When you don't, it feels off. That's a useful check, but it's only checking the yes-pile against itself.

The refusal read is different. You read the draft looking not for what's missing but for what's present that you'd never say. Not wrong, exactly. Just not yours. A phrase with an energy you don't use. A register you've never written in. A move that's plausible but would never occur to you.

Pull an AI-assisted draft and read it this way. How many moves do you find? What do they have in common?

The refusal becomes a diagnostic. When you find a sentence that isn't wrong but isn't you, that sentence is showing you part of your no-pile, making it visible for the first time, in the negative. That's useful information. It tells you something about your voice that a guide built only from kept examples may not have recorded.

Voice drift (the gradual, often unnoticed shift in a writer's prose style under the pressure of tools and templates) tends to happen through accumulation. Small compromises, kept drafts that were almost right, sentences that felt close enough. The refusal read is one way to catch it before it compounds. You're not asking if the draft sounds like you. You're asking what it said that you'd never say.

The caveat: the guide still matters

None of this means the style guide is useless. It isn't.

A good voice document can move the draft closer to your specific register. The guide may do its job for some of the work: the structural moves, the topic range, the recurring phrases you do use. Voice recovery, when a writer has drifted, can start with exactly this kind of documentation: writing down what you do and why, so you can return to it.

The argument isn't that style guides fail. It's that a guide built only from what you kept leaves out part of the work. You can add refusals explicitly. You still need to check whether the draft respects them.

What changes if you understand that is how you use the guide. You stop expecting it to be complete. You ask which pile it describes and add what's missing: phrases to avoid, moves you reject, the reason you changed a sentence. You use the guide to get the tool into the neighborhood, then read the draft differently: watching for what got through that wouldn't have.

The style guide tells the tool what you've said, and whatever refusals you've recorded. The refusal read checks what got through anyway.

That distinction is the difference between a draft that sounds like you and a piece only you could have written.

Keep your original nearby. Compare the wording with Verbaria’s free browser-based tool, then read for meaning and voice. The comparison is not an authorship detector.