Claude Projects for Writing Style: What Each Layer Actually Holds

TL;DR: Claude Projects for writing style work only as well as what you put in them. A project has three layers: instructions (applied every message), knowledge (consulted when relevant) and memory (a summary of past chats). None of those layers measures how you write. Type "professional yet friendly" into the instructions and Claude returns the average of everyone who ever typed it, because AI does not know your tone; it knows words that describe tone. Put the countable parts of your voice in the instructions instead: sentence-length range, where your qualifiers sit, the transitions you use and avoid, your punctuation habits. Each new Claude model follows that kind of instruction more precisely, so a measured Style Profile gets more useful with every release.

You created a Claude Project called "My Writing", set the instructions to "write like me: clear, warm, professional, no jargon", and uploaded a few emails to the knowledge section. The first draft opened with "I hope this finds you well." So did the fifth.

The project did exactly what you configured. The configuration just contained almost nothing about you.

What Does a Claude Project Actually Hold About Your Writing Style?

A Claude Project stores three kinds of information, and each reaches Claude differently: instructions, knowledge and memory. Understanding the reach of each layer tells you where a style rule belongs and where it will be ignored.

Anthropic's own description of what Projects are is a workspace that groups chats with shared context. Inside that workspace:

Two properties of that table matter for style.

Instructions are the only layer with guaranteed reach. Whatever sits there is present when Claude drafts your email, your report, your LinkedIn post. If a rule about how you write is not in the instructions, it applies only when Claude happens to pull it from knowledge, and Claude pulls from knowledge based on topic relevance, not on style relevance. A file of your past emails is relevant when you ask about a past email. It is not obviously relevant when you ask for a new one about a different subject.

Memory is about facts, not voice. Anthropic's article on memory in Claude Projects describes it as a way for Claude to keep context from earlier conversations in the project. It will remember that you are preparing a board deck and that the client is in Osaka. It will not derive from those chats that your sentences average eleven words and you never open with a pleasantry, because nothing in the memory layer counts those things. We covered the same distinction on the ChatGPT side in the memory versus Style Profiles comparison, and the conclusion transfers: memory stores what you talked about, and what you talked about is not how you write.

Claude also offers Styles, a separate personalisation feature with presets such as Concise and Formal, plus custom styles Claude can generate from samples you provide. A Style is a set of instructions too. The same question applies to it as to the project instructions: what did it extract from the samples, and can you read and correct it?

Why Does Claude Still Not Sound Like You After the Project Is Set Up?

Because the instructions most people write describe an effect, and Claude needs mechanics. AI does not know your tone; it knows words that describe tone, and a word that describes tone points at a register shared by thousands of writers.

Consider what Claude has when the project instructions say "warm, professional, concise". During training it saw an enormous amount of text described with those words: HR announcements, customer-support replies, consulting summaries, LinkedIn posts. The phrase now points at the statistical centre of all that text. When you invoke it, you get the centre. Your colleague with the same three adjectives in her own project gets the same centre. The two of you write nothing alike, and Claude has no way to know that, because the only thing either of you supplied was the label.

"Professional yet friendly" describes thousands of people. A voice lives in sentence length, qualifier placement, transitions and punctuation. Those four are where a reader recognises you, and none of them is an adjective:

• Sentence length is a distribution, not the word "concise". Your typical sentence, the short ones you drop to when deciding, the long ones you allow when laying out context.
• Qualifier placement is whether "probably" opens the clause or closes it, and whether you hedge at all once you have made a recommendation.
• Transitions are "so" versus "therefore" versus a bare new sentence, and whether you ever write "however".
• Punctuation is colons per email, commas per sentence, and whether asides go in parentheses or get their own sentence.

Every one of those can be counted from ten emails you already sent. A count can be written as a rule. A rule can be checked. "Was that warm enough?" has no answer; "was the mean sentence length in range?" does. We laid out the full argument in why tone adjectives are not your voice, and the Claude Project case is the same argument with a better container.

The container being better is real. Claude follows long, detailed instructions well, and project instructions have room for a full page of rules. That capacity is wasted on three adjectives.

What Should Go in Claude Project Instructions for Writing Style?

Measurements and prohibitions, stated so that Claude can verify each one against its own draft. Every line in the instructions should be something you could count in the output, and every adjective should be replaced by the mechanic it was standing in for.

Here is what the translation looks like:

Notice that the right-hand column has no adjectives. It also has a property the left-hand column lacks: when a draft misses, you can see which rule it missed and edit that one line. "Warmer" as feedback rearranges the same average. "The warmth line went in the opening, move it to the end" fixes one thing and leaves the rest alone.

Add three more sections below the mechanics:

1. Register switches. How your writing changes between a client, a peer and your manager. One line each. Claude handles conditional rules well when they are explicit.
2. Phrases you never write. This list is often more useful than the list of phrases you do. It is also where most AI-sounding drafts are caught.
3. A checklist Claude runs before it answers. "Before returning a draft: count the sentences over 26 words and split them; confirm no opening pleasantry; confirm the request is in sentence one." Claude will actually do this, and it turns your rules from aspiration into a pass.

This block is a Style Profile. Writing it by hand takes about an hour with ten emails and a word counter. MyWritingTwin.com does the measurement step from samples you already have, across more dimensions than you would count by hand, and packages the result as a Master Prompt. Your Writing Twin for Claude is that block of instructions pasted into the project, not a separate chatbot. The same block works in ChatGPT custom instructions and Gemini's saved info, which matters if your team is split across tools.

Where Do Writing Samples Belong: Instructions or Knowledge?

Rules belong in instructions, and a small set of samples belongs in knowledge as reference. Putting samples in instructions wastes the layer with guaranteed reach; putting rules in knowledge means they apply only when Claude decides they are relevant.

The instinct to upload twenty emails to project knowledge and write "match the style of the attached" in the instructions is understandable, and it does beat adjectives. Claude now has raw evidence. What it has to do next is infer, on every request, which patterns in those twenty emails are your habits and which belong to one bad Tuesday. That inference happens fresh each time, and it drifts: as a conversation lengthens, Claude's own output accumulates below the samples and pulls the draft back toward its default. We saw the same ceiling on the ChatGPT side in our look at knowledge files and personas in ChatGPT Projects, and the ceiling is the same in Claude because the cause is the same: evidence is not a specification.

The split that works:

• Instructions: the measured rules, the prohibitions, the register switches, the pre-answer checklist. Everything that must hold on every message.
• Knowledge: three to five samples, one per register you use most, labelled ("client follow-up", "internal update", "proposal cover note"). Claude reaches for these when it needs to see a shape, and the labels tell it which one.
• Memory: leave it to do its job on facts. Do not expect it to carry style.

One more reason for the split. Instructions you can read. A rule you can read is a rule you can correct when it turns out to be wrong, or when your writing changes. Samples in knowledge and the inference Claude draws from them are invisible to you. If the draft is off, you cannot see what Claude concluded from the samples, so you cannot fix the conclusion. You can only add more samples and hope.

For a walk-through of the prompt and project setup that gets Claude out of its default register in one sitting, see how to make Claude sound like you. If you want the profile to reach every chat, project and agent without pasting it each time, a connector is the durable route, described in giving a Claude connector your writing style.

Does a Newer Claude Model Make the Project Setup Unnecessary?

No, and the reason is structural. Model capability and conditioning on your data are separate layers. Anthropic improves the first with every release and cannot supply the second, because nothing about how you write is in the model.

Each new Claude follows nuanced instructions more faithfully than the last. A rule like "hedge before recommendations, never after decisions" was hit-or-miss two generations ago and is reliable now. That is a real improvement, and it improves exactly one thing: the execution of whatever instructions you provided. If the instructions are three adjectives, a better model produces a more polished version of the generic register. If the instructions are a measured profile, a better model produces a closer match to you.

The upgrade multiplies the input. Zero times a better model is still zero.

This is why "just wait for the next model" is the wrong plan for anyone who writes for a living. The next model will be better at being told. It will not be better at guessing, because there is nothing about you to guess from. Our post on why better AI models still do not know your voice works through what each release actually adds and what it structurally cannot. The short version: the profile you build today runs better on every model that follows, which makes it one of the few pieces of AI setup that appreciates.

How Do You Set Up a Claude Project for Your Writing Style, Step by Step?

Measure first, then configure. The configuration takes ten minutes once the measurements exist, and the measurements are the part most people skip.

1. Collect ten to fifteen pieces you wrote and were happy with, in the register you use most in Claude. Emails people replied to well, updates that landed, a proposal that closed. Skip anything you wrote in a hurry or heavily edited from someone else's draft.
2. Count the mechanics. Take words per sentence across the set (typical, shortest, longest), then sentences per paragraph, where the ask sits, every hedge and its position, commas per sentence, colons per piece, and the pattern of your opening and closing lines.
3. List the transitions you used and the ones you never used. The second list goes straight into the prohibitions.
4. Write the findings as rules, in the format of the right-hand column above. No adjectives. Add the register switches and the pre-answer checklist.
5. Create the project and paste the rules into project instructions. Name the project for the register it serves ("Client email", "Internal writing") if your registers differ enough to warrant separate rule sets.
6. Upload three to five labelled samples to project knowledge. One per register. Label each in the filename.
7. Leave memory alone. It will accumulate facts about your work as you use the project. It is not carrying your voice, and it does not need to.

If step two sounds tedious, that is the step MyWritingTwin.com automates: upload the samples, get the Writing DNA analysis back as a Style Profile you paste into step five. Our overview of what a writing style profile is shows what the measured dimensions look like when there are more than four of them.

How Do You Test Whether the Project Actually Knows Your Style?

Give the project the brief for an email you already sent and count the differences. Five minutes, no tools, and a number instead of a feeling.

Pick an email from last week. Write a two-line brief: who it is to, what you need, what they already know. Paste the brief into a chat inside the project and ask for the email. Put the draft next to yours and count:

• Words per sentence, shortest and longest
• Where the ask sits
• Hedges: how many, opening or closing the clause
• Commas per sentence
• The first line and the last line

If the numbers differ, the project does not know you yet, whatever Claude tells you when you ask it directly. Look at which rule was missed and edit that one line in the instructions. Run the test again on three emails in different registers. Adjective-based instructions usually pass the first test and fail the second, because the label was written for one mood and you have several. Measured rules converge across all three.

That convergence is the whole point of putting a Claude Project to work on your writing style. The project is a good container. It was always waiting for something specific to hold.

See how far your current setup sits from your real writing in about a minute with the Humanity Score test, or build a Style Profile from your own samples and paste it into the Claude Project you already use.