How to Make AI Write in Your Client’s Voice

by
How to Make AI Write in Your Client's Voice

A brand designer I work with lost a client over two paragraphs. She’d used AI to write the rationale sections of a presentation deck, the part that explains why a mark or a palette suits the brand. The client wrote back that those paragraphs read like “generic AI filler” and asked whether the strategy work was even hers. It was. Only the prose was generated. The client ended the $6,400 project at the second milestone and paid $2,800.

Nothing in those paragraphs was wrong. They just sounded like nobody in particular. Most AI copy has the same flat rhythm: every sentence lands with equal weight, and every opinion gets hedged until it says nothing.

Her prompt was the cause. She pasted the design choices into a chat window with no samples of her writing and nothing about the client. If you write for clients, or under your own name across a few channels, you’ve had a milder version of this. You ask the model to sound like someone and get back something correct and lifeless.

The fix is what the model sees before it writes. Below is the full setup: which samples to collect, the profile you build from them, where both live so the model reads them every time, the check that catches drift, and the number that tells you it’s working.

Contents

1. Why AI drafts all sound the same

This has been tested. In a 2025 study, researchers gave frontier models from OpenAI, Google, Meta and DeepSeek a handful of real writing samples from more than 400 everyday authors and asked for new pieces in each person’s style. The models did reasonably well with news articles and emails. They struggled with blogs and forum posts, where personal style is looser. In the authors’ words, the output would “often default to an average, generic tone”, and an AI detector rarely took it for human writing (never more than 55% of the time, and often under 20%).

Three more findings from that study shape everything below. Adding samples barely helped: going from two to ten examples in the prompt moved the results very little. Picking samples on the same topic as the new piece made the imitation worse, apparently because it narrowed how much of the person’s range the model could see. And the biggest gain in sounding human came from seeding the draft with a real opening fragment written by the author.

So which samples you pick matters more than how many, and a model left to itself drifts toward the middle. For a consultant with a recognizable style, or a coach whose audience follows them for how they say things, the middle is the problem. The person whose name is on the copy notices before anyone else does.

2. Collect the samples first

Before you touch a prompt, gather 10 to 20 real pieces of the person’s writing. Working writing beats polished writing: emails to their list, LinkedIn posts, a few paragraphs of a sales page they wrote themselves, transcripts from a podcast or sales call where they talked off the cuff.

Spread the samples across topics. Ten emails about ten different things will teach the model more than ten emails about the same launch. Match the formats you’ll be writing in, too. If you’re drafting their newsletter, most of the samples should be newsletters.

One of my clients, a freelance conversion copywriter who works alone, uses each client’s best-performing pieces rather than a random selection. She won’t take on a client until there are at least eight published pieces to work from, and she has turned down two prospects on that basis.

Then read the samples, looking for the patterns that survive across formats:

  • How long the sentences run. Short bursts, or long clause-heavy builds before the point lands?
  • How a paragraph opens: with the claim, or building up to it.
  • Whether they hedge or state positions flat.
  • The filler phrases they repeat. Everyone has a few.
  • Words and constructions they never use. These often tell you more than the words they do use.
  • How they get from one idea to the next. Some writers say “so” constantly. Others use no connecting words at all.

Spend 20 minutes reading the samples out loud. Rhythm is far easier to hear than to see.

3. Write the voice profile with this prompt

Once you have the samples, run this prompt. It produces a profile you can reuse in every session.

Voice extraction prompt (copy and paste):

You are a writing analyst. I'm going to give you 10 to 20 writing samples from one person. Your job is to extract a detailed voice profile that another writer (or an AI) could use to replicate their style accurately.

Analyze the samples and produce a profile covering:

1. Sentence length and rhythm: short/long ratio, how they vary sentence length for effect
2. Paragraph structure: how they open paragraphs, typical length
3. Opinion style: do they hedge or state positions directly?
4. Recurring phrases or sentence openers they use (list at least 5 real examples from the text)
5. Words or phrases they never use or actively avoid
6. Tone markers: formal/informal, warm/clinical, confident/tentative
7. How they handle transitions between ideas
8. Punctuation habits (comma-heavy, minimal punctuation, heavy use of dashes, etc.)
9. How they write calls to action or close a piece
10. Three sample sentences you would write in their voice, based purely on the patterns above

Do not summarize what they write about. Focus entirely on how they write.

Here are the samples:

[PASTE SAMPLES HERE]

Run it once per client, then edit the output yourself. The copywriter I mentioned writes every voice file herself and won’t let a client write one, because what clients say about their voice and what their copy actually does are different things. Ask clients for samples, never for a description of how they write.

4. What a voice profile looks like

Here’s what the prompt produced on a real set of samples, from Jordan M., a B2B consultant. This is the kind of document you’re aiming for.

Voice profile: Jordan M., B2B consultant

  • Sentences average 12 to 15 words. Short sentences (under 8 words) close paragraphs for emphasis.
  • Paragraphs open with the claim, then support it. Never builds to a conclusion.
  • States opinions directly. Uses “I think” rarely, and only when genuinely unsure.
  • Recurring openers: “Here’s the thing,” “Most people skip this step,” “The answer is usually simpler than it looks.”
  • Never uses: “synergy,” “it’s worth noting,” “in today’s world,” “circle back,” “as a result.”
  • Tone: confident, slightly blunt, warm when addressing the reader directly.
  • Transitions: minimal. Relies on paragraph breaks and headers rather than connecting phrases.
  • Punctuation: light on commas, heavy on short paragraphs. An occasional dash for an aside.
  • Closes with a direct next step, never a soft summary.
  • Sample sentences: “Stop optimizing the wrong thing.” / “Most consultants charge for time. The good ones charge for outcomes.” / “You don’t need a bigger audience. You need a better offer.”

Keep it to one page, and make every line specific enough that someone else could check a draft against it.

The profile on its own isn’t enough, and this is the part most guides skip. The same copywriter spent three weeks early this year describing each client’s voice in prose, and it produced nothing usable. When she gave the model twelve real examples instead, the drafts improved immediately. The samples do the work. The profile tells the model what to look for in them.

5. Where the profile lives: a Claude or ChatGPT project

The model has to see the profile and the samples every time, without you remembering to paste them in.

In Claude, create a Project for each client. Paste the profile into the project instructions and upload the samples to the project knowledge. Every chat inside the project starts with both. Free Claude accounts are limited to five projects, so past five clients you’ll need a paid plan.

In ChatGPT, use a Project the same way. The instructions go in Project settings and the samples go in as files. Free accounts get 5 files per project, Go and Plus get 25, and Pro and Business get 40, so if you’re near the cap, combine the samples into one document per format.

Skip Custom GPTs for this. Personal ChatGPT accounts (Free, Go, Plus and Pro) can no longer create new GPTs, and OpenAI is retiring custom GPTs, with standard retirement scheduled for December 11, 2026. OpenAI’s own migration path for existing GPTs is plugins, but for a voice setup a project already holds everything you need.

The instructions should load the profile and define the job. Here’s a template:

You write in the voice of [CLIENT NAME]. Their voice profile is below, and samples of their writing are in this project's files. Use the samples as your reference for rhythm and word choice. The profile tells you what to look for in them.

[PASTE VOICE PROFILE HERE]

When asked to draft:
- Match the sentence rhythm, opinion style and vocabulary in the profile and the samples
- Never use anything on the "never uses" list
- If the brief includes an opening line written by [CLIENT NAME], continue from it
- If you're unsure whether something fits their voice, flag it instead of guessing

When asked to check a draft:
- Compare it with the profile and the samples
- Quote each line that breaks from them and suggest a rewrite
- Don't rewrite the whole draft unless asked

The third drafting line earns its place. In the study above, giving the model the real opening of a piece did more for sounding human than anything else the researchers tried. If your client sends a rough first paragraph or a voice note, put their words at the top of the brief and let the model continue from them. For more on writing instructions that hold up over long sessions, see my guide to Claude system prompts.

6. The check that catches what slips through

Draft first. Then, in the same session, run a separate check. Most people skip this step, and it’s the one that makes the output usable.

The copywriter’s version is a saved prompt that returns the three phrases in her draft most unlike the client’s samples. She wrote it once and has revised it twice. Here’s a version you can use:

Compare the draft below with this person's writing samples and voice profile.

Find the three phrases in the draft that are least like how they write. For each one:
1. Quote the phrase
2. Say which pattern in the samples it breaks
3. Give a rewrite in their voice

Change nothing else.

DRAFT:
[paste the draft]

Capping it at three forces the model to pick the worst offenders instead of listing every small difference it can find. Her rule is that nothing goes to a client until the check has run and the result is logged. Running it every time is the whole discipline.

The check catches drift in one draft. To stop the same fixes coming back, every edit you make by hand should become a line in the profile. I wrote up that habit in a Bionic Business issue, Make AI Sound Like You, with a prompt that compares the model’s draft with what you actually sent and pulls out the rules hidden in your edits.

7. Onboarding a new client, step by step

Once the samples are in, setting up a new client takes about 30 minutes of your time. Collecting the samples is the slow part: the copywriter spends an afternoon per client, almost all of it gathering material.

  1. Ask the client for 10 to 20 pieces across different topics. Emails work best and LinkedIn posts are good. Transcripts from calls or podcasts capture spoken rhythm that edited writing loses.
  2. Run the extraction prompt, then edit the profile yourself until it matches what you know about the client.
  3. Create the project: profile in the instructions, samples in the files. Name it clearly so you can find it.
  4. Draft one real piece, run the check and send it. Ask the client to mark the specific lines that sound wrong.
  5. Add what they marked to the profile. One or two rounds usually closes the gap.

If you’re automating the rest of the handover too, the samples are one more thing an AI agent for client onboarding can collect at intake.

The profile changes as the client does. Styles shift, especially after a year or two of writing regularly. Revisit it every six months: pull new samples, rerun the prompt, compare the new profile with the old one and keep what changed.

With several clients this becomes an operational asset: one project per client, a current profile, and a test draft you can use as a benchmark. If you’d rather have it set up across your client list, that’s work I take on. Get in touch here.

8. The number that tells you it’s working

Track revision rounds per deliverable. It’s the cheapest number to collect, and it’s the one that moves when the voice is right.

The copywriter’s five retainer clients show what that looks like. Across 61 deliverables between June and September 2026, revision rounds fell from an average of 2.4 to 1.1, and two clients stopped asking for a voice pass at all. Her time per deliverable went from 5.2 hours to 3.1, which let her take on two more retainers in the same working week, roughly $34,800 a year with no hire and no rate change. One caveat on those figures: the revision counts come from her invoicing tool, but the time is self-tracked, and for two of the five clients the before figure was estimated from memory.

Log revision rounds in a spreadsheet next to each deliverable. If they aren’t falling after a month, look at the samples first. For choosing numbers like this, see how to tell if your AI agent is working.

9. Where this breaks

Old samples describe who the client was

If the samples are two years old, the profile describes who the client was. This is common with coaches and consultants who’ve sharpened their positioning since. Check the dates on everything you collect.

Client data needs the client’s OK

Slack messages, internal emails and call transcripts often hold things the client never meant to share with a third-party model. Ask before you paste. On Claude’s Free, Pro and Max plans, whether your chats train future models is a setting you choose (you can change it in Privacy Settings); Claude for Work and the API sit under commercial terms that setting doesn’t apply to. On personal ChatGPT plans, turn off “Improve the model for everyone” under Settings > Data controls. OpenAI says it doesn’t train on ChatGPT Business or Enterprise content by default. One catch on personal ChatGPT plans: rating a response with a thumbs up or down can send that whole conversation for training even after you’ve opted out. Some clients will want a locally run model for sensitive material. Take that request seriously.

Unreviewed copy under someone’s name

A membership business with 1,900 members let an agent answer forum questions under the founder’s name and photo. Over about eleven weeks, 14 members commented publicly that the answers didn’t sound like her, four said “AI” outright, and three cancelled citing it. One answer contradicted a strategy she had publicly reversed, because the agent was reading a knowledge base nobody had updated since March. No amount of voice tuning would have fixed that. The drafting was fine. Publishing under her name without her review was the mistake, and posting in her name has since been suspended. Nobody knows how many members noticed and said nothing.

Some voices aren’t worth copying

If the client’s voice is unclear, inconsistent or working against them, a faithful copy reproduces the problem. That’s voice development, which is a different job.

Mistakes still get through

Samples and a profile cut the error rate. They don’t remove it. Someone still reads everything before it goes out under a client’s name.

10. Where the usual advice goes wrong

Most tutorials give you one of two instructions. The first is to paste a few examples into a chat and ask for “this style”, which gets you one decent draft before the model drifts back to its defaults in the next session. The second is to write a description of the voice, which is the page of adjectives that cost the copywriter three weeks.

What works is the combination: real samples that load in every session, a short profile that tells the model what to look for in them, and a check on every draft. Set it up once per client, and each piece after that takes minutes.

What to do on Monday

Pick one client with at least eight published pieces. Pull 10 to 15 of their best-performing emails or posts from the last year, spread across topics. Run the extraction prompt and spend 15 minutes editing the profile until it reads right to you.

Create the project with the profile in the instructions and the samples in the files. Then draft one real piece, run the three-phrase check, and write down how many revision rounds it takes. That number is your baseline for the next month.

FAQs

How many writing samples do I need to build a useful voice profile?

Ten to twenty is a sensible range, but the count matters less than the spread. In the 2025 study, going from two to ten samples barely changed how closely the output matched the author, and samples that all covered one topic made it worse. Pick pieces on different subjects in the formats you’ll be writing. The copywriter in this article won’t start with a client who has fewer than eight published pieces.

Can I use this to make AI sound like me for my own content?

Yes, and it’s often easier, because you know at once when something sounds wrong. Pull your own emails, posts and long-form writing from the last six months and run the same extraction prompt. The profile will tell you things about your own style you hadn’t consciously noticed.

Should I use a Claude Project or a ChatGPT Project?

Both keep instructions and files together and load them into every chat in the project, so either works. Use the one you already pay for. Two limits worth knowing: free Claude accounts get five projects, and free ChatGPT accounts get five files per project. Custom GPTs are no longer an option for a new setup on a personal ChatGPT plan.

How do I handle a client who says “it still doesn’t sound like me” after the first draft?

Ask them to mark the specific lines that feel off rather than give an overall impression. “It doesn’t sound like me” is hard to act on. “I would never use the word ‘ensure'” or “I always open with a question” gives you a line to add to the profile. One round of specific feedback usually closes the gap.

Is it safe to paste client emails into Claude or ChatGPT?

Public posts and marketing emails carry little risk. Internal messages, client data and anything commercially sensitive need more care. Check your plan’s training setting (covered under Where this breaks), and get the client’s explicit sign-off before you use anything that wasn’t written for publication.

How often should I update the voice profile?

Every six months as a baseline, and sooner if the client has changed their positioning, launched a new offer or moved to a new audience. The usual sign is the client saying recent drafts feel off when earlier ones were fine. Between reviews, any edit you find yourself making twice belongs in the profile.

Can this work for video scripts and podcast outlines too?

Yes. The extraction prompt works on transcripts, and transcripts often show how someone builds an argument when they aren’t editing themselves. Expect more editing on loose, conversational formats: in the 2025 study, models imitated structured writing like email far better than informal blog and forum posts.

Sources

You’re subscribed. Read this week’s issue →

Get the workflows as I build them

Every week I share the agents and systems for growing your online business: the full build, the data they need, and exactly how you can run the same. Thousands of operators read Bionic Business. You should, too. No hype, no fluff. Only what's working now.

Free. One email a week. Unsubscribe whenever you like. Read this week’s issue first

Sam Woods

Written by

Sam Woods

Fractional Chief AI Officer · Founder, Stimulead and Daring Robot

Sam started with machine learning in 2016 and generative AI in 2019, writing production prompts before the practice had a name. He has advised and trained Fortune 1,000 teams across 37+ markets, and builds conversion work on proprietary datasets developed over a decade of campaigns rather than scraped. He writes Bionic Business, read weekly by thousands of subscribers.

More about Sam  ·  LinkedIn  ·  X