How to Teach AI to Write Like You

Most people try to make AI sound like them by adding instructions such as be conversational, make it warm, or write in my voice. Those instructions are not wrong. They are simply too broad to do much work.

One person's conversational is another person's painfully enthusiastic networking event. Ask for professional writing and you may receive something technically polished that sounds as though it was approved by a committee nobody wanted to join.

ChatGPT, Claude, Gemini, Grok, Copilot, and other capable AI assistants can all draft and revise text from your instructions. Whichever platform you use, the method is the same: give it evidence of what your voice actually sounds like. Your existing emails, posts, articles, and scripts are that evidence.

Voice is a pattern of decisions

Voice is often treated like decoration added after the important thinking is finished. In practice, your voice is part of the thinking. It appears in what you emphasize, how quickly you make a point, which words you avoid, when you use an example, and how directly you ask the reader to act.

That leads to a more useful interpretation: voice is judgment made visible.

Two people can communicate the same facts and produce completely different messages. One may begin with context and build carefully toward the conclusion. Another may state the conclusion in the first sentence and explain only what the reader needs next. Neither approach is universally correct. The choice reveals how each person thinks about the reader.

AI needs to see those choices. A list of adjectives cannot reveal them.

Show the pattern instead of naming it

A strong voice sample does not need to be famous, clever, or long. It needs to sound genuinely like you. Choose material you would be comfortable sending again today—not something you wrote under deadline pressure and regretted before lunch.

Look for examples that reveal several dimensions of your writing:

• How you open and close
• Whether your sentences are short, layered, or mixed
• How much context you provide
• The words and phrases you naturally use
• How you explain difficult ideas
• How directly you make recommendations or requests

Consistency matters more than volume. Ten unrelated samples from different years, audiences, and moods can give AI a blurry target. A few representative examples with a clear purpose are usually easier to interpret.

A worked example from start to finish

Consider this illustrative situation: A business owner wants to announce a live online session by email. An initial AI draft says:

We are thrilled to announce an exciting upcoming webinar packed with valuable insights and actionable strategies. Reserve your spot today and prepare to transform your approach!

Nothing is grammatically broken. It simply sounds like generic promotion.

The owner selects three short passages from previous emails that feel natural:

Quick heads-up: I'm hosting a short session next Thursday.

I won't try to cover everything. We'll focus on the few decisions that make the biggest difference.

If this is useful for where you are right now, I'd be glad to have you there.

The owner then starts a new conversation with the AI and gives it a focused request:

The passages below are examples of my natural writing voice. Identify the recurring patterns in tone, sentence length, structure, and word choice. Then rewrite the event announcement using those patterns.

Keep these facts unchanged: the session is Thursday at 1:00 p.m., it lasts 30 minutes, and it covers three ways to improve follow-up emails. Do not add benefits, urgency, or claims that I did not provide.

[PASTE SAMPLES]

The resulting draft might be:

Quick heads-up: I'm hosting a 30-minute session Thursday at 1:00 p.m. We'll look at three practical ways to improve your follow-up emails. I won't try to cover every possible tactic. The goal is to help you make the next email clearer and easier to send. If that would be useful, I'd be glad to have you there.

This is much closer, but the work is not finished. The owner notices that practical ways feels vague and that the goal is to help you sounds more explanatory than usual. Those phrases are plausible imitations, not necessarily authentic choices.

After human correction, the final version becomes:

Quick heads-up: I'm hosting a 30-minute session Thursday at 1:00 p.m. We'll work through three ways to make your follow-up emails clearer and easier to send. If that is useful for where you are right now, I'd be glad to have you there.

The improvement did not come from asking AI to be less robotic. It came from supplying evidence, protecting the facts, and judging the draft. The samples shaped the rhythm and restraint. The explicit boundaries prevented invented urgency. The final edit restored the owner's actual intent.

The same method in another context

Imagine turning a detailed article into a short social post. Instead of asking AI to make it punchy, provide two or three posts that show how you normally compress an idea. Ask it to preserve the article's central point while matching the examples' opening style, paragraph length, and degree of directness. Then check whether the shorter version still says what you believe. A recognizable cadence is not useful if the meaning has drifted.

Where voice matching goes wrong

The common misconception is that enough examples allow AI to become you. They do not. AI can reproduce visible patterns without understanding the experience, values, or responsibility behind them. It may make a stronger claim than you would, add a tidy conclusion you do not believe, or repeat a favorite phrase until it becomes a costume.

There is also a selection problem. If you provide only polished sales emails, AI may assume every message should sound promotional. If you mix formal reports with casual texts and offer no context, it may blend them into a voice you have never used anywhere.

Protect confidential information when choosing samples. Remove private names, internal details, customer data, and anything else that does not belong in the tool. Voice matching does not require exposing sensitive material.

Use the Sample–Pattern–Judgment model

A simple mental model keeps the process grounded:

1. Sample: Provide writing that genuinely represents you in the relevant context.
2. Pattern: Let AI identify and apply recurring choices rather than merely copying phrases.
3. Judgment: Review the meaning, tone, and implied promises before the writing leaves your hands.

The third step is what keeps this from becoming impersonation by autocomplete. Your name will be attached to the message, so your judgment has to remain attached to the process.

For a small test today, take one email you are proud of and one generic AI draft. Ask ChatGPT, Claude, Gemini, or whichever capable AI assistant you use to describe the observable differences, then revise the draft using the email as a style reference without changing the facts. Read the result aloud. Your ear will often catch false notes before your eyes do.

When you're ready to build and test a reusable voice setup, use the Action Guide to put this approach into practice.

You are my expert coach on: How to Teach AI to Write Like You.
Here is what I want to apply:

AI matches your voice more reliably when you provide representative examples of your writing and treat its output as a draft to judge—not a finished imitation to trust blindly.
Real writing samples give AI stronger guidance than adjectives such as friendly, professional, or conversational.
Your voice lives in recurring decisions about rhythm, emphasis, structure, vocabulary, and what you leave out.
A useful voice setup separates the facts being communicated from the style used to communicate them.

My situation: [describe your role, your goal, and what's in your way].
Walk me through it step by step, ask me one clarifying question first, then give me a specific plan I can act on today.