Brand and tone of voice: a brand book AI can actually use

Most companies and freelancers keep their tone of voice in one person's head. That person can tell that “we wouldn't say it like that,” but can't explain why — so they read every piece of text before it goes out. It works until there are ten pieces a week, until a freelancer joins, until you go on vacation. Then the website, the newsletter, and the social replies slowly start drifting apart, each in its own direction.

The fix isn't a longer document about brand values. Tone of voice can't be described with adjectives — it can be described with rules and examples. “We're friendly but professional” helps nobody, because every brand claims that about itself. “We never write ‘feel free to reach out,’ we write ‘drop us a line’” helps immediately, because you can act on it. And that's exactly the kind of document AI can use too: a yardstick it can check any piece of text against in under a minute.

This guide walks from zero to a working system. First you'll pull a style description out of writing you're already producing, without realizing you had a style. Then you'll turn it into a brand book — a living document stored where AI sees it in every conversation. Next you'll set up a consistency check anyone on the team can run. Finally you'll use the brand book to onboard freelancers and revisit it once a quarter so it doesn't calcify. Every phase comes with prompts to copy — just fill in the brackets.

A typical scenario

Klára runs a three-person marketing team at an outdoor-gear online store. She writes most of the copy herself, an outside copywriter helps occasionally, part-time staff write product descriptions, and support handles customer replies. Over a year, things drifted: the website reads formal and reserved, Instagram is bursting with exclamation points and emoji, the newsletter sounds like a press release, and support writes “Dear Customer, your claim has been registered.” From the outside it looks like four different companies.

Klára handled it the only way she knew how: reading everything before it went out. About an hour and a half a day — and the freelancer got instructions over the phone, so his first three pieces always needed a rewrite.

In one afternoon she rebuilt the whole thing. She picked ten pieces of copy she'd always liked and had a style description pulled from them — and for the first time saw, in black and white, what she actually does: short sentences, second person, concrete numbers instead of adjectives, no exclamation points, no “revolutionary.” She added values, audience, banned phrases, and twenty good/bad examples. The result was a six-page document she saved into a Project.

Her day looks different now. The freelancer got the brand book instead of a phone call, and his second piece went out without edits. The part-time staff run product descriptions through the check before sending and fix them themselves. Support has its own channel variant of the tone — same voice, just shorter and faster. Klára reads final copy for about twenty minutes a day, mostly the pieces that make a promise to a customer. Exactly how each of those steps works is the rest of this guide.

Phase 1: extracting tone from what you've already written

Don't start by writing a brand book from a blank page. Almost anyone who sits down and thinks “what kind of brand do we want to be” writes the same thing: friendly, expert, human, trustworthy. Your actual tone of voice already exists — it's scattered across the texts you wrote and were happy with. Phase 1 pulls it out of them.

The advantage of this approach is that the result is descriptive, not aspirational. It doesn't describe how you'd like to sound — it describes how you already know how to sound, and that's the only tone a team can actually maintain.

Choosing the sample: ten pieces, not fifty

Pick ten to fifteen pieces of writing you'd point to and say “that's how it should sound.” Not the most successful ones — the most typical ones. And mix channels: two articles, two newsletters, three social posts, two customer replies, one product description. If you only feed the model articles, it'll describe the style of your articles, not your brand.

Leave out anything written under deadline pressure, anything handed over by an agency, and anything you ended up having to rewrite. And leave out anything with customer personal data — replace names and addresses from support replies with placeholders before pasting anything in.

A hundred texts isn't better than ten: the model pulls an average out of a large sample, and an average is exactly what you don't want.

A prompt that describes style instead of praising it

The key is banning evaluation. Ask “what's this style like” and you'll get a compliment. You need to ask for observations that can be checked against the text.

Here are 10 pieces of copy I wrote for my brand. They're examples
of how I want to sound.

[paste the texts, each separated by a line of --- and labeled by
channel, e.g. "--- newsletter ---", "--- customer reply ---"]

Describe my writing style so that someone else could write in it
from your description. Do NOT evaluate quality, do not praise,
do not suggest improvements. Just describe what's actually in the
texts, and give a concrete example from a specific text for every
observation.

Break it into:
1. Length and rhythm: average sentence length, ratio of short to
   long, paragraph handling, where a one-sentence paragraph is used
2. Vocabulary: 20 words and phrases that repeat across the texts;
   flag separately any words a competitor probably wouldn't use
3. Address and person: formal/informal "you", we/I, how customers
   are referred to
4. Formality on a scale of 1-10, and exactly what shows it
5. Punctuation and typography: exclamation points, ellipses,
   parentheses, dashes, emoji, capitalization
6. How the texts open and close (list the actual openings and
   closings)
7. What arguments are used: numbers, stories, authority, experience
8. Humor: is it there, what kind, where does it never show up

At the end, write 5 sentences that would explain my style to
someone who's never read any of it.

It returns an analysis where you'll see, for the first time, things you do automatically — like almost never opening with a question, or a favorite construction that shows up three times in every piece. Check point 2: the model sometimes lists words that only appear once in the texts, not repeatedly. Verify each word actually recurs — otherwise you'll bake a fluke into your brand book.

Round two: what we never say

Negative definitions are more valuable in practice than positive ones. A list of banned phrases works instantly because it's binary: either the phrase is there or it isn't. “Be friendly” can't be enforced; “don't write ‘feel free to reach out’” can.

You have my style description from the previous response and my
10 texts.

Put together a list of what NEVER appears in my writing — and what
could appear there if someone else in my field had written it.

Return four lists:
1. Words and phrases that don't appear in my texts but are common
   in [industry] (marketing filler, corporate boilerplate)
2. Constructions I avoid (e.g. passive voice, sentences running
   three lines, rhetorical questions in the opening)
3. Typographic habits I don't use (exclamation points, emoji,
   all caps, ellipses)
4. Topics or claims I never resort to (comparing to competitors,
   unsupported promises, scare tactics)

For each item, write what I use instead in my texts — a concrete
example from the source. Where you don't see a replacement, say so.

This returns a list that becomes the most-used section of the whole brand book. Watch for one thing: the model sometimes bans something that's just missing by chance (emoji, say, because you happened to pick formal samples). For each item, ask yourself whether it's a real rule or a fluke of the sample — and cut the flukes.

Verifying the extraction: a blind test

A style description is useless until someone can write a text from it that actually passes. The test is simple and takes ten minutes.

Here's a description of my writing style:

[paste the description from the first prompt and the banned list
from the second]

Write three texts I've never written, based on it:
1. A short post for [channel] about [topic] (under 100 words)
2. A reply to a customer complaining about [situation]
   (under 150 words)
3. An opening paragraph for an article about [topic]

Stick to the description, not a generic idea of "good writing."
Under each text, note which rules from the description you used
and where you weren't sure, because the description doesn't
cover it.

Read the result out loud. If a piece feels off, the flaw isn't in the model — it's in the description: something important is missing or poorly captured. The last paragraph of the prompt is the most valuable part for exactly this reason: the list of places where the description goes silent is also the list of sections your brand book still needs. Run two or three rounds until the generated text stops feeling foreign.

Phase 2: the brand book as a living document

Now turn this into a document. Not a client presentation, not a sixty-page manual with gradient logo mockups — a working file a freelancer opens on Monday morning and writes correctly by Tuesday.

What has to be in it

A working brand book for writing has seven sections and fits on five to eight pages. Nobody reads more than that.

• Who we are and what we do — three paragraphs, not a mission and vision. What you sell, to whom, and how you differ from the closest alternative. The goal is for the writer to be able to picture the product.
• Values translated into writing — every value paired with a sentence: “in our writing, this means…” A value without a translation into concrete behavior is decoration.
• Audience — two to three groups, each with what they're dealing with, what they already know, what they don't understand, and the language they themselves use to talk about their problem. This section decides more about jargon than anything else.
• Tone of voice — the description from phase 1, turned into rules. Not adjectives, but sentences like “we write in second-person singular, informal.”
• Banned phrases — a “don't write / write instead” table. The most-used page in the document.
• Good/bad examples — twenty pairs, more on that below.
• Channel variants — how the newsletter differs from a support reply. Same voice, different volume.

Build a brand book for writing our brand's copy from these inputs.

Inputs:
- style description from the analysis of my texts: [paste]
- list of banned items: [paste]
- what we do: [3 sentences]
- our values: [list them]
- our audience: [describe the groups]
- channels we write for: [list them]

Requirements for the document:
- structure: who we are, values in writing, audience, tone of
  voice, banned phrases, good/bad examples, channel variants
- every rule must be checkable: someone can look at a text and
  say definitively whether the rule was followed
- no adjectives without a translation into behavior — when you
  write "friendly," it must immediately be followed by
  "which means…"
- every value gets a sentence: "in our writing, this means…"
- banned phrases as a table: don't write | write instead | why
- max 6 pages, written for someone starting here tomorrow

Where you're missing an input, write a separate line: QUESTION:
[what you need to know] — don't guess.

This returns a document skeleton plus a list of questions you never thought to ask yourself. Those questions are the most useful part of the output — they typically reveal you haven't decided whether to use informal address on social media, or how to write about competitors. Decide those yourself, not the model: AI proposes, a human approves, and that goes double for brand identity.

Good/bad examples matter more than rules

If a brand book only had one section, it should be this one. Everyone interprets a rule their own way; a pair of texts side by side can't be interpreted any other way. Always write them as pairs: the same message, once wrong, once right, plus one sentence why.

From the brand book we have tone rules and a list of banned
phrases:

[paste the rules and bans section]

Produce 20 good/bad example pairs. For each pair:
- the same message written twice
- make the WRONG version realistic — how an average copywriter
  in our field would actually write it, not a caricature
- the RIGHT version following our rules
- one sentence on which rule is being broken or followed

Cover these situations: welcome email, delay notice, claim
rejection, product description, feature announcement, reply to
a negative review, event invite, apology for a mistake, discount
offer, article closing with a call to action.

Pull the subject matter from our field: [field], product
[product].

This returns material you can paste straight into the document. Pay attention to the “wrong” variants — if they're exaggerated (something like “THE BEST DEAL EVER!!!”), they're useless, because nobody actually writes like that. Have them rewritten at a realistic level: the mistake your team actually makes is “we look forward to your reply” at the end of every email, not all caps.

Where to store the brand book

A document sitting in a year-old email attachment doesn't exist. It needs one place both people and AI can see.

The best option is Projects — persistent context that automatically attaches to every conversation inside the project. Create a project called “Copy [brand],” upload the brand book to it as a file, and write instructions for the project on how to use it. Then you no longer need to paste the brand book into every conversation — the model already has it. More on the mechanics in Projects as persistent context.

Something like this works well for the project instructions:

In this project we write and check copy for the brand [brand].
The brand book is uploaded as the file brand-book.md — it's
binding.

Rules for every response:
- Before you start writing, check tone against the "Tone of
  Voice" and "Banned Phrases" sections. Don't use a banned phrase
  even in paraphrase.
- If you're not sure how the brand would say something, don't
  guess — ask, or write two variants and flag how they differ.
- Never invent facts about the product, pricing, deadlines, or
  results. Where you're missing a figure, write [FILL IN: what].
- Match the channel variant to where the text is going. If I
  don't state a channel, ask.
- Skip opening pleasantries like "happy to help" — go straight
  into the text.

Project instructions are where the quality of every future output gets decided — it's worth investing half an hour in them. More detail in the tip on custom instructions.

If you work out of a folder on disk, Claude Cowork fits too: the brand book sits next to the drafts, and the model works with both directly — there's no copy that can drift from the original.

Make the document genuinely living: the end should carry the date of the last revision, who did it, and a list of open questions. Whenever someone hits a situation the brand book doesn't cover, they add it there. The quarterly revision (phase 5) then just works through the list that's piled up.

Phase 3: checking consistency across channels

This is where the investment pays off. The brand book stops being a document to read and becomes a yardstick anyone can check any text against in a minute — and, more importantly, fix themselves, without waiting on you.

The basic check prompt

Save this prompt in your prompt library and give it to the whole team. It's the most-used prompt in the entire system.

Check the attached text against our brand book.

The text is going on: [channel]. Target audience: [group].
Purpose of the text: [inform / sell / apologize / invite].

[paste the text]

Don't rewrite the whole text. Return a table of findings:
location in text | what it violates | which brand book rule |
suggested fix

Check in this order:
1. Banned phrases — including paraphrased or inflected forms
2. Tone: person, formality, sentence length, rhythm
3. Channel variant: does the length and structure fit [channel]?
4. Audience: are there terms [group] wouldn't understand, or
   conversely, explanations of things they already know?
5. Promises and claims: is there anything we're promising
   without backing it up?

At the end, give an overall verdict in one sentence: PASSES /
MINOR EDITS / REWRITE — and for the last two, name the biggest
problem.

This returns a list of specific spots, not general feedback. Watch two things. First, the model tends to want to improve things a rule doesn't actually cover — findings like “this sentence could be punchier” should be ignored; the check is about matching the brand book, not general quality. Second, don't let the model rewrite the text outright. Fix everything at once and you'll get a text that technically meets the rules but has lost your point. Fix it manually based on the findings — it takes five extra minutes and the text stays yours.

Channel variants: same voice, different volume

One tone doesn't mean one text everywhere. A newsletter can handle paragraphs, a chat reply can't. A professional-network post lives or dies on the first sentence, a product description on its structure. That's why the brand book needs a channel table — and having it drafted for you takes ten minutes.

We have this tone of voice:

[paste the Tone of Voice section from the brand book]

We write for these channels: [list them, e.g. website, newsletter,
LinkedIn, Instagram, support replies, sales emails, product
descriptions].

For each channel, write a channel variant:
- length: typical text length and paragraph length
- structure: how it opens, how it closes, what never belongs
  there
- what from our tone gets amplified and what gets toned down
  (and why)
- 3 typical mistakes we tend to make on this channel
- one sample sentence in our tone showing what an opening on
  this channel would look like

The tone must stay recognizably the same — don't invent new
personalities for individual channels. The difference is in
volume, not voice.

This returns a table you add to the brand book as the final section. The last paragraph of the prompt matters: without it, the model happily suggests “on Instagram we're more playful and use emoji,” even if emoji is on your banned list. Require every channel variant to be consistent with the base tone, not an exception to it.

Auditing existing content

By the time the brand book exists, you already have years of writing produced without it. Rewriting everything isn't worth it — knowing where the drift is worst, and fixing the spots that get the most traffic, is.

Here are 15 of our existing texts from different channels, plus
our brand book.

[paste the texts, each labeled with channel and date]

Do a consistency audit:
1. Rank the texts by how far they drift from the brand book
   (most drifted first), with a score from 1-10 and the main
   problem in one sentence for each
2. Find patterns: which mistakes repeat across texts, not
   one-off exceptions
3. Split mistakes into: writer error × gaps in the brand book
   (spots where the document is silent and everyone handles it
   their own way)
4. Suggest a fix order by impact-to-effort ratio

Be specific on point 3 — I want to know exactly what needs to
be added to the document so this stops happening.

This returns a priority list and — more valuable — a map of the holes in your own document. Point 3 typically reveals that the brand book says nothing about writing numbers, addressing people in mass emails, or apologies. Fix the document before you fix the texts.

The check should be run by the writer before sending, not by the boss afterward. You only read the texts that make a promise to a customer — pricing, deadlines, guarantees. That's not about tone, it's about commitment, and sending, promises, and anything with legal weight are always approved by a human.

Phase 4: onboarding freelancers and agencies

The most expensive part of working with an outside copywriter isn't the fee — it's the three rounds of edits before they get how you sound. The brand book cuts that to one.

A brief instead of a phone call

Give the freelancer two things: the brand book and a brief for a specific piece. The brief can be produced with a prompt that forces you not to leave anything important out.

Prepare a brief for a freelance copywriter. You have our brand
book.

Text I need: [type of text], channel [channel], length
[word count], deadline [date].
Topic and facts: [list what the text needs to cover]
Goal: [what the reader should do or know after reading]
Audience: [group from the brand book]

Put together a brief that includes:
- a clear split: what's fixed vs. where the writer has freedom
- which brand book sections matter most for this text
- 3 examples of our existing texts that are the closest stylistic
  match (pick the type based on the brand book, I'll add the
  specific pieces)
- what "done" means: 5 criteria the text will be accepted against
- what the writer should avoid — pull out the banned phrases
  from the brand book most likely to show up in this type of
  text

The brief has to make sense to someone who's never heard of us.

This returns a brief that's ready to send. The “what done means” section is the one that saves rounds of edits — when a writer has acceptance criteria up front, they won't hand in a text you then have to argue about.

A trial task with immediate feedback

Before you hand a freelancer paid work, give them one short piece and run it through the phase-3 check. Feedback stops being “something's off” and becomes a list of specific spots tied to a rule. The difference in how fast they learn is dramatic: a writer who gets “you broke the exclamation-point rule, see page 3” remembers it. A writer who gets “this sounds too marketing-y” keeps guessing.

The same goes inside the team: the check prompt is the best tool for onboarding a new hire — the first week they run texts through it, by the second they're anticipating the mistakes. Ties into onboarding a new hire through a Project.

Agencies and subcontractors

At an agency, the person writing the copy is someone you don't even know. Send the brand book as an attachment with the order, and agree upfront that the agency runs the check before delivery, not you afterward — with the check prompt in hand, there's nothing left to argue about.

We have a brand book and work with an agency. Put together a
one-page "How to write for us" document to attach to the order.

It must include:
- the 10 most important rules of our tone, ranked by how often
  they get broken
- a table of banned phrases (only the most common, max 15 rows)
- 5 good/bad example pairs
- the check the writer should run before delivery
- what counts as grounds for sending a text back

Write it for a professional, not a beginner — concise, no
explaining the obvious. It needs to fit on one A4 page.

This returns a condensed version of the brand book that works as an attachment. It doesn't replace the full document — the one-pager is the ticket in, the brand book is the reference manual.

Phase 5: evolving tone and the quarterly revision

A brand that sounds exactly the same in 2026 as it did in 2021 either has an unusually strong identity, or it's stagnant. Tone shifts along with the product, the audience, and whoever's writing for the brand. The question isn't whether it changes, it's whether it changes on purpose.

What changes and what doesn't

Split the brand book into two layers. The core — values, the relationship with the customer, boundaries, what we'll never do — almost never changes; changing it is a leadership decision, not a document revision. The surface — specific vocabulary, channel variants, examples, the banned-phrase list — changes regularly, because it ages: a phrase that was fresh two years ago is a cliché today, and a channel you weren't on back then now carries half your communication.

The quarterly revision, step by step

Set aside ninety minutes once a quarter. The process has four steps: the open questions you've collected, checking the document against reality, updating it, and logging the change.

We're doing the quarterly brand book revision. I'm attaching:
- the current brand book
- 10 texts we published over the past quarter (chosen as
  representative, not our best work)
- open questions we've collected over the quarter: [list them]

Go through it and return:
1. Where reality diverges from the document — rules we're
   systematically breaking. For each one, say whether the mistake
   is in the texts, or the rule is unsustainable and should
   change.
2. Which good/bad examples are now outdated (referring to
   products, channels, or situations we no longer deal with)
3. Phrases that have crept into our texts and should be added
   to the banned list — pull them from the attached texts
4. Answers to the open questions: for each, propose a rule
   written in the same style as the rest of the document
5. What's redundant in the document: rules nobody has ever
   broken or needed, because they're obvious

Don't rewrite the document, return a list of proposed changes
with reasoning.

This returns a set of proposed changes for you to decide on. Point 1 is the most valuable and the most uncomfortable: it forces you to decide whether to enforce a rule or admit it doesn't hold anymore. Either is better than a document where half the rules only exist on paper. Point 5 then trims the document — and a shorter document actually gets read.

Versioning: make changes visible

Write the brand book in markdown, not a slide deck: two text files can be diffed line by line, so after a revision you can see exactly what changed. Anyone using Git gets history for free; everyone else can just save dated versions. After every revision, send the team and freelancers a summary — not the whole document, just what's new.

Compare the old and new versions of the brand book:

OLD:
[paste]

NEW:
[paste]

Write a message for the team and outside collaborators:
- what changed, ranked by impact on everyday writing
- one sentence why, per change
- what it means in practice: "from now on, instead of X we write Y"
- what stays the same (just a list, so people don't have to read
  the whole document)

Keep it under 300 words, written in our tone per the new version.

This returns a message you can send out. The last line of the prompt is a small test: if a message about the brand book isn't written according to the brand book, something's off.

A different case is a deliberate shift: you're changing audiences, moving into a new segment, merging two brands. There, you can't start from existing texts, because those describe who you were — write down who you want to be, generate examples, and test them on people who know you. And expect it to take months: until your highest-traffic pages and most common reply templates are rewritten, the old brand lives on.

Common mistakes

• A brand book full of adjectives. “We're authentic, bold, and human” helps nobody — you can't write a single sentence from a document like that. Every adjective needs a “which means, in our writing…” sentence and an example.
• Writing the brand book from a blank page instead of extracting it. An aspirational tone nobody on the team actually knows how to write doesn't get followed. Start from what you can already write, and shift it gradually.
• A document nobody can find. A brand book sitting in an old email attachment doesn't exist. It needs one place — ideally a Project — where people and AI both see it.
• Letting AI rewrite a text straight from the brand book. You get a text that follows the rules and has lost the point. The check returns findings, a human fixes them.
• Confusing consistency with uniformity. A support reply shouldn't look like an article. Channel variants are part of the brand book, not a violation of it.
• Freezing the document. An unrevised brand book ages faster than you'd think after a year — especially the banned-phrase list, since clichés turn over.
• Checking tone and missing content. A text can be perfectly on tone and still promise a deadline you won't hit. Facts, prices, and commitments are always checked by a human.

The best tools

• Projects (claude.ai) — home for the brand book: a file plus project instructions, so the model has the rules in every conversation without pasting anything in.
• Claude Cowork — working over a folder where the brand book sits next to drafts in progress; a revision propagates instantly and there's no outdated copy floating around.
• A prompt library — the check prompt, the freelancer brief, and the revision prompt saved so the whole team uses them, not just you; more in the prompt library tip.
• Markdown and versioning — the brand book as a text file: you can diff version against version and show exactly what changed.
• Examples instead of description — whenever a rule is hard to describe, replace it with a pair of examples; the technique is laid out in the tip on format examples.

What you get out of it

• Time: checking a text before sending drops from “the boss reads it whenever they have time” to a minute, done by the writer. For a team publishing ten texts a week, that's hours a week — and, more importantly, the waiting disappears.
• Money: a freelancer who gets the brand book and a brief delivers a usable text in round one or two instead of round three. The saved rounds of edits are the most expensive part of working with outside help.
• Peace of mind: you stop being the bottleneck. Texts can go out while you're on vacation, because the yardstick is written down, not in your head.
• Quality: consistency across channels is something customers notice without being able to name it. A brand that sounds the same everywhere reads as someone who knows what they're doing.

Pro tip

The fastest way to stress-test a brand book is to have it write a text about something your brand has never dealt with before — an apology for an outage, a price-increase announcement, a response to public criticism. That's where you have no existing samples to fall back on, and it reveals whether the document is really a guide or just a description of what already exists. If the model flounders on a text like that or produces a generic corporate response, your brand book is missing a section on crisis situations — and you write that one calmly, in advance, not on the day you need it.

And one final rule: the brand book is a yardstick, not an authority. When someone on the team writes a text that breaks a rule and is obviously better for it, the mistake isn't in the text — it's in the rule. Log it as an open question and settle it at the next revision. A document that never bends for a genuinely good text stops being a tool and becomes an obstacle.