Write With AI While Keeping Your Own Voice: A Practical Method

Write With AI While Keeping Your Own Voice: A Practical Method

To write with AI while keeping your own voice, do not ask it to invent your voice. Show it. Give it passages you wrote yourself, describe in short rules what is typical for you and what you avoid, and review every draft with a fixed checklist before publishing. AI speeds up the first draft, but the voice remains your decision.

Key takeaways

• AI imitates what it sees better than what it reads in a description. Real samples of your writing are worth more than adjectives like direct or warm.
• Short, checkable style rules, covering what to do and what to avoid, work better than vague requests.
• A good prompt combines four pieces: context, samples, rules and task.
• Review is part of the method: you cut what is not yours, check the facts and adjust the tone before publishing.
• Save what worked in a reusable style guide so you do not start from zero with every text.

Why does AI-written text often sound generic?

Without a reference, a language model tends to pick the most common way of saying each thing. The result is correct and neutral, and for that very reason hard to recognize. It is the text that opens with a long greeting, uses filler phrases and closes with a courtesy line nobody reads.

Your voice is made of repeated choices: sentence length, the words you prefer, how you open and close, what you would never say. Those choices are hard to capture in adjectives. If you ask for a professional and friendly tone, the model returns the average version of professional and friendly, which is everyone's version.

The fix is to replace description with evidence. When the model sees how you actually write, it has a concrete pattern to follow. What follows is a working method based on hands-on implementation and editorial common sense, not a guarantee that the output will match your own writing. That is why human review remains mandatory.

There is also a quality question that has nothing to do with the tool. Google Search Central's guidance on helpful content asks for material made for people, with reliable information and real value. Text that only rearranges commonplaces does not meet that bar, whether you write it with AI or without.

Step 1: Which samples of your own writing should you collect?

Start with a small selection of your texts. Quality matters more than quantity. Anthropic's documentation on context engineering recommends curating a small set of varied, representative examples instead of piling up rules for every possible situation. It also helps you avoid wasting space in the context window.

Use these criteria to choose:

• Pick only texts you would publish today without changes. If something has aged badly, leave it out.
• Vary the format: a short email, a post, a reply to a customer, a longer passage.
• As a practical starting point, not a rule, begin with three to five samples and add more only if the results are unstable.
• Remove names, phone numbers, amounts and any customer or employee data before pasting the text. This applies under Brazil's LGPD and under plain common sense.
• If you like, include one sample labeled as what should not happen, to show the contrast.

One detail that helps: organize the samples in separate, labeled blocks and say what each one was for. The model tells a sales message from an internal notice more reliably when that is written down.

Step 2: How do you turn what you notice into style rules?

Read your samples as if someone else wrote them and note the patterns. Ask: how does the text open? What is the typical sentence length? Which words repeat? What is missing, such as exclamation marks, emojis or jargon? How does it end?

Then convert each observation into a rule you can check. A good rule is one where another person can look at the text and answer yes or no. Here is an illustrative style guide:

• Short sentences, one idea each. At most one long sentence per paragraph.
• Direct opening, with no long greeting and no obvious scene-setting.
• Address the reader as you. No formal titles, no slang.
• Avoid filler phrases such as in today's world, it is worth noting and it is important to highlight.
• No emojis and no strings of exclamation marks.
• Vocabulary: customer, not user; quote, not commercial proposal.
• Close with a clear next step, never with feel free to reach out.

Rules like be authentic or use a human tone do not help, because you cannot verify them. Also prefer a few important rules to a huge list. When the list grows too long, rules start competing with each other and some get ignored. If you already write instructions for other purposes, the logic is similar to writing instructions for an AI support agent: a few clear instructions with explicit limits.

Step 3: How do you build the prompt for the AI?

With samples and rules in hand, build the prompt in four parts, always in the same order. A repeated structure makes it easier to compare results and see what changed.

• Context: who you are, who the text is for and what it should achieve. One or two sentences are enough.
• Samples: the chosen passages, labeled by type, with the instruction that they are a style reference and not content to copy.
• Rules: the short style guide, including what to avoid.
• Task: what to write now, with the facts that must be included. Supply the facts yourself, as bullet points, so the AI does not have to invent anything.

Two practices improve the result noticeably. The first is to ask for a short draft before the full text, to validate the tone early. The second is to say explicitly what the AI should do when information is missing: ask, not fill the gap. This reduces invented claims.

When the text comes out different from what you expected, avoid vague complaints. Point to the passage and name the cause, for example: this opening is too long, my pattern is to go straight to the point. Then carry that correction into the style guide so it applies to the next texts.

How do you review the draft so it sounds like you?

Review is where the voice gets confirmed. Read the text aloud, or at least in your head at your own pace. If you stumble, the sentence probably is not yours. Use a fixed checklist, which you can extend with the checklist for reviewing AI-generated text:

• Facts: did you check every number, name, date and promise? AI is not a source of truth.
• Opening and closing: do they follow your pattern, or did they slide back to the generic template?
• Vocabulary: do any words appear that you would never use?
• Rhythm: do the sentences vary the way they do in your samples, or are they all the same length?
• Filler: are there paragraphs that say little and could go without loss?
• Commitments: does the text promise something your company cannot deliver?
• Personal data: is there any information that should not be there?

If the same problem shows up in several drafts, the fix is not to keep editing by hand but to add a rule to the guide. Over time, review gets shorter because the guide absorbs your adjustments. The reverse also holds: if a rule never changes the result, remove it.

Illustrative scenario: an accounting firm's newsletter

Imagine, as an illustration, a partner at an accounting firm who writes a monthly newsletter for clients. She picks four earlier editions she likes, strips out names and amounts, and labels each one. She writes the month's topics as bullet points, along with deadlines she has already verified.

In the first draft, the AI opens with Dear readers and closes with We hope this was helpful. Neither appears in her editions. She does not rewrite the whole text. She adds two rules to the guide: open directly with the main topic, and close with the action the client needs to take. By the second draft, the structure already resembles her earlier editions.

During review, she finds a sentence that suggests broader tax guidance than the firm would put in writing. She cuts it, checks the deadlines against the official source and publishes. The gain here is not that the text comes out effortlessly. It is that the time spent structuring and drafting is smaller than the time spent deciding what is true and what is her position.

What not to hand to AI, and how to apply a method

Not every text should be drafted by AI. Sensitive messages, such as a reply to a customer in a dispute, a termination notice or anything with legal consequences, call for your own writing or, at the very least, a careful review. It is worth reading about what not to delegate to AI in business and setting a company policy for employee ChatGPT use, mainly to know which texts and data may be pasted into external tools.

The method we use in AI projects fits writing well, at a small scale:

• Map: list the types of text you produce and identify which are repetitive and low risk. Record how you write them today.
• Limit: define which data may go in, which types of text the AI may draft, and that nothing goes out without your approval.
• Pilot: choose a single type of text, use the workflow for a few weeks and compare, round by round, whether you would publish it with minor edits. Record what you had to correct.
• Hand over: document the style guide and the samples so another person on the team can use the same standard, and decide whether to expand, adjust or stop.

If the idea is to expand to a team, the style guide becomes a shared asset and should be treated like one: with an owner, a version and periodic review. Tools that let you save reusable instructions help, but what matters is the content of the guide, not the tool. In public-facing replies, such as online reviews, the same care for natural tone applies, as the guide to responding to Google reviews with AI without sounding like a robot shows.

Frequently asked questions

Do I need to train or fine-tune a model for it to write like me? In most writing cases, no. Samples and rules in the prompt itself are usually enough to start, and they are easier to update. Training or fine-tuning a model takes more data and maintenance, and only makes sense with large volume and a stable pattern.

Can I paste customer texts as style samples? Not without care. Remove names, contact details and any information that identifies people, and check the tool's rules on data use. Prefer your own texts, already free of personal data, and consult Brazil's data protection authority (ANPD) if you have doubts about data handling.

Does the AI keep my voice in long texts? It tends to drift from the pattern as the text grows. Write in sections, repeat the main rules at each stage and pay special attention to reviewing the ending, where drift usually shows up.

Do I need to disclose that I used AI to write? It depends on the context, the platform and the rules of your industry, and there is no single answer. The central point is that you take responsibility for what you publish, check the facts and decide what goes out under your name.

Sources

• Google Search Central: conteúdo útil, confiável e feito para pessoas
• Anthropic: Effective context engineering for AI agents
• AWS: comparando recuperação de contexto e fine-tuning