Writing with AI and still sounding like yourself

The draft arrives in nine seconds. It is organised, it is grammatical, it has a tidy triplet in the second paragraph, and it could have been written about any business by anybody. You read it, you cannot point at a single wrong sentence, and you send it. Then nothing happens. No reply, no forward, no argument. Nothing was wrong with the writing, and that was the problem.

This guide is for solo operators, freelancers and small-team owners who write their own newsletters, proposals, client updates and sales pages, and who have started drafting with a model. It covers how to brief one with your actual writing rather than a description of it, how to split the job so you never rubber-stamp a finished draft, which patterns to cut on the way out, and what you are handing a vendor when you paste your best paragraphs into a chat box. It is not a guide to beating AI detectors, which are not reliable enough to design around [5]. It is also not the right document for copy where the wording itself carries legal risk, such as regulated financial or medical claims, where the question is what you are allowed to say and the answer comes from a lawyer rather than an editor.

The model supplies competence and nothing else about you

A language model produces the most plausible continuation of the text in front of it. Given a topic and no other context, the most plausible continuation is the middle of everything ever written on that topic, and the middle has no clients, no numbers and no position. What an LLM actually is covers the mechanics. The practical consequence is narrower than it sounds: the model is not withholding your voice. It was never given it.

Anthropic’s prompting documentation puts the fix in one sentence: “Examples are one of the most reliable ways to steer Claude’s output format, tone, and structure” [4]. Notice what that sentence groups together. Tone sits next to format and structure, treated as the same class of problem and solved the same way, by showing rather than telling. The same page offers a test worth stealing for every brief you write: “Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.” [4] A colleague handed the instruction “write this in my voice” would have to ask which voice, and would be right to.

Show it your writing instead of describing it

Adjectives do not transfer. “Warm but professional, direct, not corporate” describes several thousand mutually incompatible documents, and asked to average them the model will do exactly that. A sample transfers, because a sample is not a description of a voice, it is an instance of one.

Anthropic recommends 3 to 5 examples for best results, and asks that they be relevant, mirroring your actual use case closely; diverse, varied enough that the model does not pick up unintended patterns; and structured, wrapped in tags so they can be told apart from instructions [4]. Translated into writing terms, relevant means the same genre as the thing you are about to write, so client emails for a client email and not your best blog post. Diverse means not five pieces from the same week about the same product, or the model will learn that topic instead of that voice. Structured means labelling them clearly as samples, so the model imitates them rather than answering them.

Two further rules come from the material rather than the docs. Choose pieces from a good day rather than an average one, because the model will reproduce whatever level you hand it and your median output is not the thing you want cloned. And use whole pieces rather than paragraphs, because the parts of a voice that readers actually register, when you run long and when you stop early, live at the length of a page.

Then park the samples somewhere you will not rebuild them every Tuesday. Claude handles this through styles, which you create by uploading a file from your device or pasting the text directly, in pdf, doc or txt, alongside the four presets of Normal, Concise, Formal and Explanatory [1]. You can switch between styles at any time during a conversation [1], which matters more than it reads: the voice for a client apology is not the voice for a landing page, and keeping two styles costs less than re-explaining yourself twice a week.

ChatGPT keeps the equivalent in custom instructions, available on all plans on web, desktop, iOS and Android [2]. The field holds up to 1,500 characters on Free and Go, and up to 5,000 on Plus, Pro, Enterprise, Business and Education [2]. That is room for rules and not for samples, so treat the field as your style sheet and keep the writing itself in a project or at the top of the session. One behaviour to know before you rely on it: “updates to your instructions are reflected only in future conversations” [2], so fixing a rule halfway through an argument with a draft will not retroactively fix that draft.

Gemini’s version is a Gem. You name it and “write instructions for it to follow”, and you can attach files, which Google describes as a way “to give your Gem more context or to reference specific docs in your chats with the Gem” [3]. Google’s own advice for the instruction field is to “provide details about your goals, desired behaviors, and preferred format” [3], which is a fair summary of what a style sheet is for.

The tool matters less than the asset. The asset is one file containing 3 to 5 complete pieces of your writing that you would be content to have imitated, plus a short list of rules you actually follow. Build it once.

Split the job so you never approve a whole draft at once

The request that reliably produces the most generic result is “write me the post”. Every gap in that instruction gets filled with an average choice, and there are hundreds of gaps.

Split it. First ask for substance and only substance: the points worth making, the objections a sceptical reader would raise, the examples that might carry the argument. Then argue with the list. Cut the points you would not have made yourself, add the two you would, and reorder it so the strongest claim is not buried in position four. A list is cheap to fight with. A finished draft is not, because a finished draft arrives with the authority of being finished, and the path of least resistance is to accept its shape and tidy the sentences.

Only then draft, one section at a time, from the points that survived. This keeps you in the editor’s chair rather than the subscriber’s, and it has a second effect that is easier to feel than to explain: at section length you still notice when a paragraph is saying nothing, and at article length you stop noticing.

Say what you want rather than what you are avoiding. Anthropic is explicit about this asymmetry, recommending that instead of “Do not use markdown in your response” you try “Your response should be composed of smoothly flowing prose paragraphs” [4]. It applies directly to voice. “Do not sound like AI” gives the model no target. “Short sentences, one idea each, no summary sentence at the end of a paragraph, name the client rather than saying a client” is a target. The general technique is in prompting fundamentals.

Edit for the tells, and ignore the detectors

Machine drafts have habits, and the list is short enough to hold in your head. Inflated significance, where a routine product change marks a pivotal moment and a small fix underscores the importance of something. The rule of three, applied until every sentence arrives as a neat triplet. Negative parallelism, the “it is not just X, it is Y” construction that sounds like an insight while asserting nothing. Vague authority, the sentence that leans on research or on experts with no research and no expert attached. A suspicious evenness, every paragraph the same length and the same shape, because nothing in the process ever got interested and ran long. And the summary sentence at the end of a paragraph that restates the paragraph, which survives more editing passes than any of the others because it feels like good structure.

Deleting them is half the job. The other half is what goes in the hole. Ceremony out, specific in: the customer’s actual name, the figure from your own invoicing, the thing you got wrong last year, the sentence that runs long because you got interested. Those are the parts a model cannot supply, for the straightforward reason that it does not have them, and they are the parts readers use to decide whether anyone is home.

What you are not doing here is dodging a scanner. OpenAI has been asked directly whether AI detectors work and answered: “In short, not in our experience. Our research into detectors didn’t show them to be reliable enough given that educators could be making judgments about students with potentially lasting consequences.” [5] Their own attempt at a classifier “labeled human-written text like Shakespeare and the Declaration of Independence as AI-generated” [5]. They note that such tools “sometimes suggest that human-written content was generated by AI”, with a disproportionate effect on students writing in English as a second language and on writing that is “particularly formulaic or concise” [5]. And they add the part that closes the subject: “Even if these tools could accurately identify AI-generated content (which they cannot), students can make small edits to evade detection.” [5] Edit for the tells because a human reader registers them, not because a piece of software might.

Some writing stays yours, and someone has to own the rest

Keep a category of writing that never goes near a model, and not for sentimental reasons. Apologies. Condolences. Anything resolving a conflict. Anything where the person on the other end is quietly asking whether a human actually considered them. People absorb a clumsy sentence from a person without difficulty. They do not absorb discovering that their complaint was answered by a template with feelings. The rule that holds up in practice is that AI helps you write about things, and you write to people.

There is now a legal version of the same instinct in one jurisdiction, and it repays reading even if it does not apply to you. Article 50 of the EU AI Act, which applies from 2 August 2026 [6], states that “Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated” [6]. The scope is narrower than the internet’s summaries of it suggest. Your proposal, your client update and your sales page are not text published to inform the public on matters of public interest.

The instructive part is the exemption. The obligation “shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content” [6]. Read that as a description of a workflow rather than as a loophole. The test for whether AI-assisted text needs a warning label is whether a named human read it and is answerable for it. That is a reasonable test for a newsletter no regulator will ever open. This is orientation and not legal advice; if you publish news or public-interest material into the EU, the applicability question belongs with a lawyer.

Your voice file is data you are handing to a vendor

The samples are your own writing, so the ownership question there is easy. The material you paste alongside them is often not yours: the client’s brief, the half-finished case study, the numbers out of somebody else’s business. The terms differ by product and by setting, and you cannot tell which regime you are under by looking at the interface.

Anthropic’s position for the consumer plans, Claude Free, Pro and Max, is a choice you make: “We will use your chats and coding sessions (including to improve our models) if: 1. You choose to allow us to use your chats and coding sessions to improve Claude” [7]. Incognito chats sit outside it entirely, and “are not used to improve Claude, even if you have enabled Model Improvement in your Privacy Settings” [7]. Submitting feedback is a separate act with a longer tail, since the related conversation can be stored “in our secured back-end for up to 5 years” [7].

OpenAI’s Data Controls “allow you to choose whether your conversations help improve our models”, and with the setting off, “Your conversations will still appear in your chat history but won’t be used to train ChatGPT” [8]. Temporary Chats are deleted after 30 days, are not used to train the models, and do not create memories [8]. Team, Enterprise and Edu plans carry additional data controls [8]. One detail specific to this guide sits on the custom instructions page rather than the privacy page: “Information from your use of custom instructions will also be used to improve model performance”, though “Your instructions won’t be shared with shared link viewers” [2].

None of this is an argument against pasting your writing into a chat box. It is an argument for knowing, once, in writing, which account you draft in, what its setting says about training, and what you have quietly agreed to paste there on a client’s behalf. Ten minutes with the settings page beats re-litigating the question every time you open a draft.

What still goes wrong

Voice samples teach the model your surface, not your judgment. It will learn your sentence length, your habit of starting with the objection, your refusal to use the word solution, and it will attach all of that to a point you do not actually hold. The draft that sounds most like you is the one whose errors are hardest to catch, because the usual alarm, the sentence that does not sound like something you would write, has been switched off. Verification gets more important as the imitation improves, not less, and the routine for it is in the verification habit.

The samples also decay. If you feed back drafts that were themselves AI-assisted, each round moves your reference material a little closer to the average, and after a year of it the file no longer contains your voice. It contains the model’s impression of your voice from last spring. Refresh it from things you wrote by hand, and keep the pieces you are proudest of out of the rotation entirely so there is always an uncontaminated copy.

Human review has the same weakness wherever it appears, including in the exemption quoted above [6]. At two pieces a week it is a genuine editorial pass. At forty it becomes a click, and the checklist gets satisfied by someone who scrolled. Nothing in this guide fixes that, because it is a volume problem rather than a tool problem. And the hardest limit is the one nobody advertises: if you have not written enough by hand to have a voice yet, there is nothing to put in the file. The model cannot give you one. It can only return, at speed and at scale, whatever you were already able to show it.