Claude Tone: What the Settings Control and What They Cannot

You picked the Formal style, added "warm but direct" to a Project, and Claude still writes like a competent stranger. The settings worked exactly as designed. They just were never designed to know you.

TL;DR: Claude tone is controlled by three stacking layers: Styles in the app, Project instructions, and the system prompt on the API. Every one of them shifts the register. None of them contains a single fact about how you write. A tone label ("professional yet friendly", "warm but direct") describes a category with thousands of members, and Claude produces that category's average. Your actual voice lives in sentence-length range, qualifier placement, transitions and punctuation. Put those measurements into the same layers and the tone settings start doing what you thought they did.

What Does "Claude Tone" Actually Refer To?

Three separate controls, each in a different place, all of which change the register Claude writes in. People searching for "Claude tone" usually mean the first one and end up needing the third.

The Styles feature sits closest to what most people mean by tone. Pick Formal and Claude drops contractions and lengthens sentences. Pick Concise and it trims. The custom option asks you to describe the style you want or paste a sample, then Claude writes an instruction for itself and applies it going forward.

Project instructions do the same job with more room. Anthropic's own Projects documentation describes them as context that persists across every conversation inside the Project: who you are, what you are working on, how you want responses shaped.

The system prompt is the developer version. Anthropic's guidance is to use it for role and context, and it is the layer that every app built on Claude, including the Claude app itself, ultimately talks through.

All three accept the same input: text describing how to write. That shared input is where the problem starts.

Why Does Claude Still Sound Generic After You Set the Tone?

Because a tone setting selects a category, and Claude writes the statistical centre of that category. The setting did its job. The job was smaller than you assumed.

Take the Formal style. Formal is a register shared by lawyers, HR departments, academic reviewers and the person writing a complaint to their landlord. Claude has read all of them. When you select Formal, it moves toward the average formal writer in its training data. That writer has a sentence length, a hedging habit and a way of opening paragraphs, and none of them are yours.

The custom Style option looks like it fixes this, and it moves the needle slightly. Type "warm but direct, no fluff, short paragraphs" and Claude generates a paragraph of instruction from that. The instruction is still made of adjectives. "Warm" and "direct" are labels for buckets, and a bucket has a middle.

We tested this on the tone-adjective problem for ChatGPT in Describe My Writing Tone to AI? Adjectives Are Not Your Voice, and Claude behaves the same way for the same reason. A description of tone contains no measurement. Claude does not know your tone. What Claude knows is the words people use to describe tone, and it maps those words to the average writer who would use them.

Here is the gap in concrete terms. Two managers both describe themselves as "professional yet friendly":

• Manager A writes 11-word sentences on average, opens emails with the ask, uses "probably" as a softener at the end of the sentence, never uses exclamation marks, signs off with initials.
• Manager B writes 22-word sentences, opens with a line of context, softens with "I think" at the start, uses one exclamation mark per email, signs off with a full name and title.

Same adjectives. Two voices with nothing in common except the label. Feed the label to Claude and both managers get the same email back, because the label was the only thing Claude received.

What Is a Voice Made Of, If Not Tone?

Measurable habits that repeat across everything you write in a given context. They are countable, and because they are countable, an instruction layer can follow them.

The dimensions that separate one writer from another, in rough order of how much they matter:

1. Sentence-length range. Not the average alone but the spread. Some writers sit between 8 and 15 words and rarely leave. Others swing from 4 to 40 inside one paragraph.
2. Qualifier placement. Which hedges you use ("probably", "I think", "it seems", "roughly") and whether they open the sentence, close it, or sit next to the verb.
3. Transitions. How you move between points. Some people use "So," and "Also," as paragraph starters. Some never start a paragraph with a connector at all.
4. Punctuation density. Commas per sentence, colon use, parentheses, whether you ever use a semicolon, whether you use em-dashes (and if so, how many per page).
5. Openers and closers per context. How your first sentence to a client differs from your first sentence to your team.
6. The absences. Words and moves you never make. These are as identifying as the ones you do.

None of these are tone. All of them are what a reader recognises when they say "that sounds like you". The full breakdown of how these are extracted is in How Style Extraction Works.

A Style Profile is simply those measurements written down, per context, in a form an instruction layer can execute. Writing DNA is the measured layer under it. Neither one contains an adjective doing structural work.

How Do You Set Claude's Tone So It Actually Sounds Like You?

Put measurements, not adjectives, into whichever layer you use, and put them in the most specific layer available. Here is the order of operations.

Step 1: Measure before you describe. Take five to ten pieces you wrote yourself in one context (say, emails to clients). Count average sentence length and note the shortest and longest. List every hedge word you used and where it sat. Write down how each piece opened and closed. Note anything you never did. Ten minutes with a word counter gets you a rough version; a Writing Twin for Claude, ChatGPT and Gemini does it precisely from the same samples.

Step 2: Write the instruction as rules, not vibes. Compare:

Adjective version: "Write in a warm, professional, direct tone."

Measured version: "Average sentence 12 words, range 6 to 20. Open with the ask, one line of context after. Soften with 'probably' at sentence end, never 'I think' at the start. No exclamation marks. No bullet lists in email. Close with first name only."

Claude can execute the second one. It can only approximate the first.

Step 3: Pick the layer by scope.

• Writing for yourself across many topics: a custom Style, pasted as the measured instruction rather than described.
• One recurring context (client emails, board updates, a newsletter): a Project with the measured instruction in its custom instructions. Claude Projects for Writing Style covers what each Project layer should hold.
• Building a tool for a team: the system prompt, with one profile per writer or per house voice.

Step 4: Keep contexts separate. Your Slack voice and your board-memo voice are two profiles, not one profile with a "be more formal" toggle. If you write in more than one language, that multiplies again, because register does not translate; keigo, tu/vous and Sie/du each carry their own structural rules.

Step 5: Test on a known email. Pick a message you already sent. Give Claude the same brief and compare the output line by line against what you actually wrote. Sentence length, opener, hedges, closer. Where they differ, the profile is missing a rule. Add it.

Does a Better Claude Model Make Tone Settings Unnecessary?

No. A better model makes a measured profile more valuable and a vague one more disappointing, for the same reason.

Model capability and knowledge of your writing are separate layers. Anthropic improves the first with every release: better instruction following, better handling of nuance, fewer defaults leaking through. Anthropic cannot supply the second, because the second is made of your emails, and your emails are not in the model.

So when a newer Claude arrives and follows instructions more faithfully, two things happen at once. If your instruction is "warm but direct", the model produces a more precise average of warm-but-direct writers, and the output feels more polished and no more like you. If your instruction is a measured profile, the model executes each rule more accurately, and the output moves closer to you than the previous model managed.

This is the pattern we described in Does AI Know My Writing Style? Why Better Models Still Don't. The upgrade cycle rewards people who have already measured their voice and does nothing for people who have only named it.

What Is the Difference Between a Claude Style and a Style Profile?

A Claude Style is a preset that lives in one app on one account. A Style Profile is a portable measurement that runs inside a Style, a Project, a system prompt, or any other AI tool.

The two are not rivals. The best place to put a Style Profile inside the Claude app is a custom Style or a Project instruction. The Profile is the content; the Style is the slot.

What to Do This Week

Open the Project or Style you currently use for writing. Read the instruction. Count the adjectives. If the instruction is mostly adjectives, Claude has been guessing, and the output you have been editing is the average of a category.

Replace the adjectives with measurements from five pieces you wrote yourself. If you would rather not count sentences by hand, MyWritingTwin.com builds a Writing Twin for Claude, ChatGPT and Gemini from your samples and hands you the measured profile to paste into any of the three layers above. Either way, the test is the same: give Claude a brief for an email you already sent, and see whether it comes back as you.