Call · 30 min
MarketingMattia Esposito12 November 202614 min read

Your brand's tone of voice in ChatGPT and Claude. A prompt lasts one chat; a voice document, uploaded in the right place, lasts for all of them.

Ask ChatGPT to write “in our brand style” and you get the style of every other brand. Here is the voice document to upload, where to put it in each assistant, and a before and after on the same copy.

In brief

To make ChatGPT or Claude write in your brand’s tone of voice you need a written document, uploaded to a project, and a human read before anything is published. A prompt on its own covers one conversation, and memory is managed by the assistant, not by you.

Where it goes has changed. OpenAI retires custom GPTs on 11 December 2026 and points to projects for writing work; for Claude, Anthropic documents instructions, projects and skills.

The document has eight parts: who is speaking to whom, four voice traits, the tone on Nielsen Norman Group’s four dimensions, tone by channel, rules stated positively, constructions to cut, approved texts, and where to upload it.

You can download it below, in Word and PDF, with no email required.

Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.

The brand tone of voice document, to download

A two-page document to fill in once and upload to the assistant you write with. Each part sits next to an example for a made-up company, an industrial shelving manufacturer, plus an empty column for yours. The rules are written positively, because OpenAI recommends concrete instructions over long lists of prohibitions. The file is in Italian; the structure works in any language.

The brand tone of voice document: who speaks to whom, the four voice traits, tone on the four dimensions and tone by channel, with a worked example
FileContentsLink
Wordto fill in and upload

The eight parts with the worked example and a column for your company. This is the file to upload to the project.

tono-di-voce-brand-chatgpt-documento-voce.docx

PDFto read or print

The same document, to keep beside you when you edit a draft.

tono-di-voce-brand-chatgpt-documento-voce.pdf

PartWhat to writeWhere it comes from
1. Who speaks to whom

Who you are, who reads, what they should do next, formal or informal address.

Instructions with a role (OpenAI)

2. The voice in four traits

Four positive statements, each explained in two lines.

A constant voice (Mailchimp)

3. The four dimensions

A score from 1 to 5 for formal, serious, respectful and matter-of-fact.

Nielsen Norman Group

4. Tone by channel

Website, email, complaint, LinkedIn, quote: one example sentence for each.

A tone that shifts (Mailchimp)

5. What to do

Five rules, written positively.

Positive instructions (OpenAI)

6. What to cut

Seven constructions to look for when editing, with the rewrite.

Wikipedia, signs of AI writing

7. Approved texts

Three real texts and one not to copy, with the reasons.

Acceptable and unacceptable examples (OpenAI)

8. Where to upload it

A ChatGPT project, a Claude project or skill, a Gemini Gem.

Official help pages

When AI writes a text for me, I read it through before publishing and look for the same things every time. Long dashes opening an aside. Sentences built on “it isn’t this, it’s that”. Strings of three adjectives that sound good and say nothing. Sentences that deny one thing instead of stating another. I take them out one by one, and what remains is nearly always shorter.

On 14 October I counted the path my site had taken: 106 versions of the home page, 22 of them in a single day, five headlines in 67 days. The first landing page had been generated by Claude, and the first person to tear it apart did so just as easily. In the same post I wrote a rule I still use when I edit: “Good copy doesn’t announce what it’s going to show you next.” (the full post, in Italian).

For a small business the lesson lies in the order of the parts. The positive rules come first and guide the drafting; the constructions to cut come afterwards and serve the edit. A document made only of prohibitions produces copy that avoids all of them and still doesn’t sound like you.

What I counted, in my copy and other people’s

Five headlines in 67 days. From a line anybody could have written to one with a verb that none of 23 Italian profiles used.

106 versions of the home page. Twenty-two in a single day, starting from a landing page generated with Claude.

8, 9, 2, 1, 0. The number of times Spotify writes “AI” in its five press releases about its DJ: the story is further down.

1,153 comments read for a positioning project. “Discipline” appeared three times: buyers’ words come before tone.

Brand tone of voice: what it is, with examples

Brand tone of voice is the way a brand sounds in writing: the words it picks and how formal it is. Mailchimp’s public style guide separates two things: “You have the same voice all the time, but your tone changes.” The voice is the identity and stays put; the tone adapts to the channel and to the reader’s mood.

Nielsen Norman Group, the American usability research firm, reduced tone to four dimensions, each a spectrum: formal or casual, serious or funny, respectful or irreverent, matter-of-fact or enthusiastic. It tested them with a survey of 50 people in the United States, using texts identical except for tone: the differences people perceived were real but small, between half a point and a point on a five-point scale.

The sample is small and American only, so the four dimensions are useful in practice more than in statistics. They give the document a shared vocabulary: “2 out of 5 on formality” means the same to an AI assistant and to a new colleague. For the visual and strategic side of the brand, colours and positioning, see the guide on how to create a brand book.

I found an example of shifting tone at Spotify. On 13 October I counted how many times it writes “AI” in its five press releases about its DJ: 8, 9, 2, 1, 0. In 2023 it was an AI guide who knows you, in 2025 a friend who gets you; in May, introducing the Italian voice, Spotify describes a personality. The feature is the same, and the way it is told has moved towards a person (the post, with sources, in Italian).

How to make ChatGPT write in your brand voice

In ChatGPT the brand voice belongs in a project: the document uploaded as a file, the short rules in the project instructions. OpenAI says so on its help page: “Project instructions apply only within that project and override your global custom instructions.” The company voice stays there, and the other chats on the account are left alone.

The account’s custom instructions are the other route, limited to 1,500 characters on the Free and Go plans and 5,000 on Plus, Pro, Business, Enterprise and Edu. They apply in every chat, shopping lists included. OpenAI warns that their content may be used to improve its models unless you opt out, so customer data stays out of the document.

Two routes many guides still recommend no longer hold. Personal accounts can no longer create custom GPTs, and existing ones are retired on 11 December 2026; migration turns them into plugins, with the instructions becoming a skill. Memory, for its part, decides on its own what to keep, and OpenAI writes that it does not retain every detail of every conversation.

WhereWhat it does for the brand voiceWorth knowing
ChatGPT, project

Instructions that apply only in the project and override the account’s; uploaded files act as sources.

A good answer can be saved to the project’s sources

ChatGPT, custom instructions

They apply in every chat, including chats already open.

1,500 characters on Free and Go, 5,000 on other plans

ChatGPT, custom GPT

Can no longer be created on personal accounts.

Retired on 11 December 2026

ChatGPT, memory

Managed by the assistant, so no place to keep rules.

Does not retain every detail

Claude, account instructions

They apply in every conversation.

Better for personal preferences than for a brand

Claude, project

Instructions and uploaded documents, always present in the project’s chats.

Up to five projects on the Free plan

Claude, skill

Switches on when relevant, even outside the project.

Code execution must be turned on

Gemini, Gem

Instructions plus files from your computer, Drive or NotebookLM.

A Drive file stays up to date; Gems do not work in Gemini Live

This guide is about the tone of written copy. ChatGPT’s spoken voice, the one you pick for voice conversations, is a separate setting and has nothing to do with your brand.

Brand voice in Claude: projects and skills

Claude takes the brand voice from a project with the document uploaded, or from a skill that applies it wherever it is needed. Anthropic’s page on personalisation lists three levels: account instructions, project instructions and skills, which among other things adjust the tone and format of answers.

The project matches ChatGPT’s: a workspace with its own chats and its own knowledge base, available on the free plan too, up to five projects. Anthropic’s own example of project instructions is a more formal tone. Anthropic also notes that a new version of projects is in beta, starting from Claude Code, so menu labels may change in the coming weeks.

A skill works differently. Skills are sets of instructions, plain text is enough, which Claude loads when they are relevant, in any conversation; they need code execution turned on and are available on every plan. Anthropic’s first example of a custom skill is applying brand style guidelines to documents and presentations. For anyone writing across several channels it is the handiest place.

A word on a feature many guides still mention: the old help page on Claude’s styles no longer loads, and the official personalisation page no longer names them. If you see them in the interface, use them; for a document meant to last, projects and skills are the documented routes. In Gemini the equivalent is the Gem, with the document linked from Drive.

How the brand voice gets into an AI assistant: three real texts, the voice document, the project or Gem where it is uploaded, the draft, the edit, and the approved text going back into the document as an example Real textsthree, already approved The documenteight parts Where it goesproject, skill, Gem The draftwritten by AI The editconstructions to cut The textapproved your archiveNN/g, voice and toneOpenAI, Anthropic, Google ChatGPT, ClaudeWikipediaa person the approved text goes back into the document as an example
The path of a text written with AI in the brand voice. The document sits in the middle: written once, uploaded where you work, and enriched with every approved text.

A before and after on the same copy

The before and after below is for the made-up shelving manufacturer from the document. On the left, sentences of the kind an assistant tends to write without guidance, which I wrote for this guide using the signs Wikipedia collects; on the right, the same idea with the voice document, and in the last column the part of the document that changed it.

BeforeAfterPart of the document
In today’s industrial landscape, warehouse safety plays a crucial role.

A shelf that gives way shuts the warehouse for days.

6: empty opening and emphasis; 2: answer before introducing yourself

We don’t sell mere shelving, we create tailor-made solutions.

We design the shelving around the load you give us, pallet by pallet.

6: the sentence that denies; 2: we give the numbers

Quality, reliability and innovation for over thirty years.

Every installation ends with a signed inspection report.

6: three adjectives in a row; 5: a proof for every promise

Our team of passionate experts is always by your side.

The site survey is done by the technician who then oversees the installation.

2: we talk like people on site

Don’t hesitate to contact us for any information.

Send us the warehouse floor plan: the quote arrives within three working days.

5: a single action at the end

Let’s discover together why you should choose us.

Cut.

6: announces what it will show next

The “after” asks the company for facts: the three days for a quote and the inspection report. The voice document does not invent them, it goes looking for them. If a fact is missing, the sentence goes, and every promise left has to be backed up like any other advertising copy, with the check in the guide to copywriting for small business.

Signs of AI writing, and why to cut them

The constructions typical of language models are collected on a public Wikipedia page, “Signs of AI writing”, updated almost daily. It lists sentences stressing “importance and legacy”, negative parallelisms, the rule of three, formulaic conclusions and a vocabulary that keeps coming back. The list I use when I edit belongs to the same family.

ConstructionHow to rewrite itWhat the source says
The long dash opening an aside

A comma, or two sentences.

Wikipedia moves it to the historical indicators, 2022 to September 2026: recent models avoid it

“It isn’t a product, it’s a solution”

Only the second half, with a fact.

Negative parallelism (Wikipedia)

Three adjectives or three sentences in a row

Just one, with the proof.

Rule of three (Wikipedia)

The sentence that denies

The same idea, stated positively.

Positive language (Mailchimp); positive instructions (OpenAI)

“Plays a crucial role”

The fact that makes it matter.

Emphasis on importance and legacy (Wikipedia)

“Here’s what you’ll discover”

Say the thing.

The 14 October post, above

Opening on “today’s landscape”

Start from the customer’s question.

Formulaic openings and conclusions (Wikipedia)

Wikipedia itself warns that a single sign proves little: “This sign is most useful when taken in combination with other indicators, not by itself.” The long dash is the example, and the reason the list is for rewriting, never for judging who wrote the text.

Research offers a small business two useful findings. In six experiments with 4,600 participants, published in PNAS in 2023, people could not recognise self-presentations written by AI, and leaned on misleading cues such as first-person pronouns. A 2024 study in PNAS Nexus, with 4,976 participants in the United States and the United Kingdom, found that an “AI-generated” label lowers trust in news headlines, even true or human-written ones, because readers assume nobody checked them.

Neither study measures small-business advertising copy, so treat them as pointers. For anyone selling, the takeaway is the human edit: it makes a text yours, and it is what readers assume is missing when they think “AI”. The transparency duties for generated content are in the guide to the AI Act for marketing. As for machine checks, a 2023 study of 14 detectors found them “neither accurate nor reliable”.

Brand tone of voice for a small business: formal address, and what stays out of the document

If you sell into Italy or most of Europe, your voice document needs a line English guides skip: formal or informal address. The tone sources, Nielsen Norman Group and Mailchimp, are American, and English has no tu or Lei. An Italian business selling to other businesses usually picks per channel, often Lei in a first email and tu on social media, and the document has to say so.

The second point is what you put in it. The document ends up on a provider’s systems, and ChatGPT’s custom instructions may be used to train models unless you opt out. Keep out customer names, confidential prices and personal data; the approved texts in part 7 should come from copy already published.

The third is raw material. Tone is easier to write from your customers’ own words: reviews and questions put to salespeople. Where to find them is in the guide to understanding customers before you write; how to use them in the opening lines of a text, in the guides on hooks and headlines and on how to write a Meta ad. The About us page, where tone shows more than anywhere else, has its own guide with a template for a small business.

The other marketing guides for small businesses, from brand to outreach, are gathered in the small business marketing index; copywriting built on a market analysis is on the copy and sales materials page.

Questions and answers

What is a brand’s tone of voice?

The way a company sounds in its writing: the words it picks and how formal or casual it is. The voice stays the same everywhere; the tone shifts with the channel and with the reader’s state of mind, as in a reply to a complaint.

Nielsen Norman Group measures it on four dimensions: formal or casual, serious or funny, respectful or irreverent, matter-of-fact or enthusiastic.

How do I get ChatGPT to write in my brand voice?

Write a voice document, with traits, rules stated positively, constructions to cut and three approved texts, and upload it to a project. Project instructions apply only inside the project and override the account’s, so the brand voice leaves your other chats alone. Every text gets a human read before it goes out.

Can I still create a custom GPT with my brand voice?

Not on a personal account: Free, Go, Plus and Pro can no longer create or publish new GPTs. OpenAI has set the retirement of custom GPTs for 11 December 2026 and points to projects for writing work. If you already have one, you can migrate it or move to a project.

How do I give Claude my brand’s tone of voice?

Anthropic’s help page lists three ways: account instructions, which apply everywhere; project instructions, with the voice document uploaded; and a skill, which Claude switches on when it is relevant, even outside the project. Anthropic’s first example of a custom skill is applying brand style guidelines.

Can an AI detector tell me whether my copy sounds machine-written?

No. A 2023 study of 14 tools found them neither accurate nor reliable, and Wikipedia, which collects the signs of generated text, advises against relying on detectors. You get more from an edit with the list of constructions to cut, and from asking whether you would sign the text yourself.

Notes on sources

  1. OpenAI, Help Center: ChatGPT Custom Instructions (character limits and use to improve models), Projects in ChatGPT (source of the quotation), Memory in ChatGPT. Read on 12 November 2026.
  2. OpenAI, Help Center: Creating and editing GPTs (positive instructions, acceptable and unacceptable examples, end of creation on personal accounts) and Custom GPT retirement and migration FAQ, updated on 10 November 2026: retirement on 11 December 2026, migration into plugins, projects as the alternative. Dates may change: check them again.
  3. Anthropic, Claude Help Center: What are projects?, Understanding Claude’s personalization features and What are skills?, updated between 23 and 30 October 2026. Read on 12 November 2026.
  4. Google, Gemini Help, Use Gems in Gemini Apps. Read on 12 November 2026; plans and file limits are not covered on this page.
  5. Kate Moran, Nielsen Norman Group, The Four Dimensions of Tone of Voice, 2016, updated on 30 January 2024: the four dimensions and the survey of 50 American respondents.
  6. Mailchimp, Content Style Guide, Voice and Tone, licensed CC BY-NC 4.0: quoted for its line on voice and tone, neither reproduced nor adapted in the downloadable document.
  7. Wikipedia, Signs of AI writing, a WikiProject AI Cleanup page, licensed CC BY-SA 4.0, as of 12 November 2026; source of the quotation. The page changes often.
  8. M. Jakesch, J. T. Hancock, M. Naaman, Human heuristics for AI-generated language are flawed, PNAS 120 (11), 2023; S. Altay, F. Gilardi, People are skeptical of headlines labeled as AI-generated, PNAS Nexus 3 (10), 2024. Abstracts read via Europe PMC.
  9. D. Weber-Wulff and others, Testing of detection tools for AI-generated text, International Journal for Educational Integrity 19, 2023: 14 tools; source of the words quoted. Today’s tools differ from the ones tested.
  10. The first-person counts come from Mattia Esposito’s public LinkedIn posts of 13, 14 and 21 October 2026, linked in the text; quotations from them are translated. The before-and-after texts and the example company are made up for this guide.
·Who does it for you

A voice document takes an afternoon to write. Knowing which voice that market needs is the work that comes first. Positioning is built starting from the market.

Mattia Esposito, who wrote this guide, also works with Itria on positioning: in-depth market analysis to find where competitors are exposed and take that ground. He did it on his own site first. In August, asked who had founded itria.io, ChatGPT couldn’t find him. From 11 October to 7 November 2026 itria.io pages appeared 13,803 times in Google and Bing results and 3,818 times in AI answers: Google’s AI Overviews and AI Mode, and Microsoft Copilot.