14 min read

How to Write AI Prompts That Sound Like You: A Guide for Founders Who Hate Generic Copy

By Anastasia, founder of The Scale Atelier

AIPromptsBusiness Management
Cover image for “How to Write AI Prompts That Sound Like You: A Guide for Founders Who Hate Generic Copy”

A coach I work with sent me two versions of the same welcome email. One she had written herself in twenty minutes, tired, between calls. One an AI tool had produced in four seconds from the prompt "write a warm welcome email for new coaching clients." She asked which I preferred. It was not close. Hers had a sentence about the fact that the first session is usually the one where people apologize for their own goals. The AI version had "I am thrilled to embark on this transformative journey with you." Generic AI copy is not a sign that AI cannot write. It is a sign that nobody told it who was speaking. The pull toward sameness is measurable. A 2026 study in Nature Human Behaviour analyzed over 880,000 texts and found that when large language models polish and rewrite writing, they reduce the variance in writing complexity by 21% to 50%, "emphasizing conformity over individuality." Left to its defaults, the tool averages you out.

This article is about not being averaged out. It covers why AI copy sounds like everyone else's, what "sounding like you" actually consists of, how to write a voice brief the tool can follow, why examples beat adjectives every time, which parts of writing you should keep doing by hand, and a one-hour setup you can do this week. By the end you will have a reusable brief, a small set of examples, and a review habit that lets you use AI for drafts without your readers noticing anything except that you got back to them faster.

Why does AI copy sound like everyone else's?

The short answer: because the model's default voice is the average of everything it has read, and a vague prompt asks for the default. "Write a warm email" is a request for the median warm email on the internet.

The research on this is consistent and a little unsettling. The Nature Human Behaviour study found that the flattening holds "across different LLMs, prompts and contexts," and that models "amplify patterns associated with dominant characteristics while suppressing others." In plain words: whatever is most common in the training data gets louder, and whatever is distinctive about you gets quieter. A separate controlled experiment published at the 2025 CHI conference, with 118 participants from India and the United States, found that AI writing suggestions led the Indian participants to adopt Western writing styles, "altering not just what is written but also how it is written." The tool did not just change their words. It changed their rhythm.

Here is the reframe I offer founders who feel personally insulted by their AI drafts. The tool is not ignoring your voice. It has never met your voice. Anthropic's prompting guidance suggests thinking of the model "as a brilliant but new employee who lacks context on your norms and workflows." Nobody would hand a new hire a laptop on day one and say "write a warm welcome email" and expect it to sound like the founder. You would show her three emails you had sent, tell her what you never say, and explain who the reader is. The generic draft is the predictable result of a generic brief. That is good news, because briefs can be fixed in an hour.

What does "sounding like you" actually consist of?

The short answer: a small number of repeatable decisions, most of which you make without noticing. Voice feels mysterious from the inside and looks like a checklist from the outside.

Before you can brief a tool on your voice, you have to be able to describe it, and most founders have never tried. Read five things you wrote that felt right and look for the decisions. How long are your sentences, and where do you deliberately use a very short one? Do you open with a scene, a question, or the point? Which words do you use that your competitors do not, and which words would you never use even under duress? What do you refuse to promise? Do you write "you" or "clients"? Do you use humor, and what kind? A founder I know realized her entire voice came down to four rules: short first sentence, one concrete detail per paragraph, no exclamation marks, and always name the thing the reader is probably worried about. Four rules. The AI could follow four rules. It could not follow "warm but professional."

This is where brand clarity comes first and intelligent systems come second. If your ideal client is "women 25 to 45 who want to grow their business," no prompt on earth will make the copy sound specific, because the reader is not specific. That is not an ICP. That is a continent. Ask "which kind?" three times until you can picture one woman reading the email at her kitchen table, and your voice sharpens on its own, with or without a tool. Consistency is worth the effort commercially as well. A 2019 Lucidpress study of over 200 organizations found that consistent branding can increase revenue by up to 33%, and that 81% of companies still deal with off-brand content. An AI tool with no brief is an off-brand content machine. The same tool with a clear brief is a consistency machine. If the clarity is the part that is missing, that is what the positioning work at the studio exists for, and it comes before any prompt.

How do you write a voice brief the AI can actually follow?

The short answer: write it once, in plain language, the way you would brief a capable new hire, and reuse it every time. The brief is the system. The individual prompts are just requests to it.

Both major AI companies publish guidance on this, and it is more practical than the word "prompt engineering" suggests. Anthropic's rule is to be clear and direct: "the more precisely you explain what you want, the better the result." Its test is one I now use on every brief: show the prompt to a colleague with minimal context and ask her to follow it. "If they'd be confused, Claude will be too." It also recommends explaining the reason behind an instruction, because the model "is smart enough to generalize from the explanation." So instead of "no exclamation marks," write "no exclamation marks, because my readers are senior and tired and read enthusiasm as pressure." The second version teaches. OpenAI's prompt engineering guide describes the same layering: standing instructions that cover "tone, goals, and examples of correct responses," which take priority over the individual request underneath.

A voice brief that works has six parts, and none of them is an adjective.

  1. Who is speaking. One or two sentences on what you do and for whom. "I am a pricing strategist for women running consultancies of two to eight people."
  2. Who is reading. The one woman at the kitchen table. Her situation, what she is worried about, what she has already tried.
  3. The rules. Four to eight concrete decisions from the previous section: sentence length, openings, punctuation, how you address her.
  4. The banned list. Words and phrases you never use. Mine includes "journey," "unlock," "game-changer," and any sentence that begins with "In today's."
  5. The phrases you actually say. Three to five lines lifted from your own writing or your calls. These do more than any rule.
  6. What you never promise. Leads, revenue, hours saved, transformation. If your brand does not promise results, your prompt has to say so, or the tool will promise them for you.

Save it as a document. Paste it at the top of every writing request, or set it as the standing instruction in whichever tool you use. It is not a prompt. It is a job description that happens to be read by software.

Why are examples more powerful than adjectives?

The short answer: because "warm" means a thousand things and one of your emails means exactly one. Every prompting guide I trust says the same thing, and my experience agrees.

Anthropic calls examples "one of the most reliable ways to steer" a model's output format, tone, and structure, and recommends three to five of them. Its advice on choosing them is worth following precisely. Make them relevant, meaning they "mirror your actual use case closely." Make them diverse, so they "cover edge cases and vary enough" that the model does not copy an accidental pattern, such as every example starting with the same word. OpenAI describes the same technique, few-shot learning, as steering the model toward a task "by including a handful of input/output examples in the prompt, rather than fine-tuning the model," and adds: "try to show a diverse range of possible inputs with the desired outputs." Neither of these is complicated. Both are usually skipped.

In practice, this means collecting a small library. Two or three emails you sent that clients replied warmly to. One social post that sounded like you at your best. One paragraph from a proposal that closed. Label each one: "welcome email, new client, after a hard discovery call." Then, when you ask for a draft, include the two examples closest to the task. The tool will match the length, the rhythm, the amount of detail and the level of formality far more accurately from two real examples than from a paragraph of description. A caution from the research: the examples need to be yours. The CHI study above shows how quickly a writer drifts toward the model's defaults when the suggestions are doing the leading. Your examples are how you make the model do the following instead.

Which parts of writing should you keep doing by hand?

The short answer: the idea, the specific detail, and anything a client will read as a relationship rather than a message. The tool drafts the logistics. You write the connection.

There is a limit to what better prompting can do, and it is important to know where it is. A 2025 study in Computers in Human Behavior: Artificial Humans compared human-written essays with GPT-generated ones and found that each additional human-written essay contributed more new ideas than each additional AI essay, and that the gap widened at scale. The authors tried to fix it from the model side and could not: enhancing the model's creative diversity "through parameter- or prompt-modifications does not mitigate the diversity gap." Read that as a job description. The new idea has to come from you. The observation that first-session clients apologize for their own goals came from twenty years of sitting across from people, and no prompt would have produced it. Your job in the writing process is to supply that sentence. The tool's job is to build the paragraph around it.

The second boundary is the one I hold most firmly. Automate the logistics around relationships, never the relationships. The welcome email, the reminder, the FAQ reply, the follow-up after a webinar: draft those with the brief and the examples, review, send. The message to a client who has just had a difficult month, the note after a disagreement about scope, the congratulations when she finally raises her prices: those you write yourself, slowly, badly if necessary. Warmth is the asset AI makes scarcer, which is exactly why it is worth more when it is real. And every draft, even the routine ones, gets read before it goes out. AI output is a starting point to review, never a finished answer. Deliberate beats fast. I wrote about what that looks like across a whole business in beyond the grind.

How do you set this up in one hour this week?

The short answer: collect, describe, brief, test, and keep a "not me" list. One hour, once, and then a few minutes each time you write.

  • Collect (15 minutes). Find five pieces of your own writing that sound like you. Emails count. Voice notes transcribed count. Label each with the situation.
  • Describe (15 minutes). Read them and write down the decisions: sentence length, openings, banned words, the phrases you actually say, what you never promise.
  • Brief (15 minutes). Turn the decisions into the six-part voice brief. Save it somewhere you can paste from. It lives in your own accounts, in plain language, so anyone you hire can use it too.
  • Test (10 minutes). Ask for a draft of one real email using the brief and two examples. Then run the colleague test: would a smart stranger, reading only your brief, know what you meant? Fix the confusing lines.
  • Keep a "not me" list (5 minutes, ongoing). Every time a draft includes a phrase you would never say, add it to the banned list. The brief improves each week, and the drafts drift back toward you.

That is the whole setup. Both guides recommend checking your prompts against real results rather than trusting them once. OpenAI's version is to build "tests and evaluation suites that measure prompt behavior." Yours can be a folder of before-and-after drafts and an honest read. If you would like help turning the brief into a system that drafts your routine messages while you are with clients, that is the shape of the AI systems work at the studio.

Founder to founder: the goal was never to make the machine sound human. It was to make sure that when your reader hears from you, she hears you. The tool is capable of that, on one condition: that you have told it who you are, shown it what that sounds like, and kept the sentences that only you could write. Write the brief this week. The first draft you get back will be the moment you stop hating AI copy, because for the first time it will not be generic. It will be yours, with the typing done.

Frequently Asked Questions

Why does AI writing sound so generic?

Because a vague prompt asks for the average of everything the model has read. Research in Nature Human Behaviour found that models reduce the variety in writing complexity by 21% to 50% when they rewrite text, so without a specific brief the output drifts toward the middle.

How do I make ChatGPT or Claude write in my voice?

Give it a plain-language voice brief and two or three real examples of your own writing. Describe the reader, the rules, the banned words and what you never promise, then include the examples closest to the task. Adjectives like "warm" and "professional" do very little on their own.

How many examples should I include in a prompt?

Three to five, according to Anthropic's guidance, and a handful according to OpenAI's. Choose examples that closely match the kind of writing you want and that vary enough that the model does not copy an accidental pattern.

Should I tell the AI what not to write?

Yes, and tell it why. A banned list of words and phrases is one of the highest-value parts of a voice brief, and explaining the reason behind each rule helps the model apply it to cases you did not list.

Can AI replace writing to my clients entirely?

No, and it should not. Use it for the logistics: welcome emails, reminders, routine replies, all reviewed before sending. Messages that carry the relationship, such as repair, celebration and scope conversations, stay handwritten.

How often should I update my voice brief?

A little every week, whenever a draft contains something you would never say. Add the phrase to the banned list and, if a draft gets it right, save that draft as a new example.

The Scale Atelier builds AI systems for women-led coaching, consulting and expertise businesses, including the voice briefs and drafting systems that keep routine messages sounding like you while you are with clients. If you would like to know where to start, book a free 20-minute audit: no pitch, just a prioritized plan that is yours to keep either way.

Free tool

Put a number on the hours.

The Time-Drain Calculator turns the time you spend on invoicing, follow-ups and admin into what it actually costs you over a year.

Open the Time-Drain Calculator

Two minutes · no email, no sign-up

Want a hand with this?

Start with a free intro call — twenty minutes, no pitch, no pressure. You leave with a clear next step whether or not we work together.