How do I get ChatGPT to write in my voice?
You get your voice out of ChatGPT by giving it three to five samples of your own writing, asking it to derive a style sheet of measurable rules from them, correcting the rules it got wrong, and then storing that sheet where every chat can see it. Naming an adjective does not work: "friendly and conversational" describes a category that millions of pages already occupy, so the model hands you the average of that category, which is the sound you were trying to escape.
The method below takes about twenty minutes once. After that the rules live in your settings and every chat starts closer to you. It works the same way in Claude and Gemini, because nothing in it depends on a quirk of one model.
Why does asking for a friendly, conversational tone not work?
Because an adjective is a label for a large group of writing, and the model can only answer it by averaging that group. Millions of pages have been described as friendly and conversational. What they have in common is the surface: contractions, second person, short paragraphs, a question in the opening line. You will get all four of those and none of you.
Your voice is not a label. It is a set of habits that can be counted. Two writers can both be "warm and direct" and share almost no habits at all. One writes 12 word sentences and never uses a list. The other writes 30 word sentences with three clauses and bullets everything. The adjective cannot tell them apart, so neither can the model.
| What you probably said | Why it produces the average | The countable version |
|---|---|---|
| "Conversational" | Names a genre with millions of members. | "Average sentence 14 words. No sentence over 30. Use contractions everywhere." |
| "Punchy" | Every writer thinks their own writing is punchy. | "Open every section with a sentence under 8 words. One idea per paragraph, 2 to 4 sentences." |
| "Professional" | The safest word in the language, so it returns the safest prose. | "No exclamation marks. No second person. Cite a number or a source in every claim." |
| "Not too formal" | A negative with no target, so the model guesses a midpoint. | "Contractions yes. Jargon allowed: churn, runway, ARR. Banned: utilise, leverage, robust." |
| "Like me" | The model has never read you. | Three pasted samples plus the rules derived from them. |
| "Like Morning Brew" | Actually works, if the reference is public and specific. | Keep it, and add the two or three habits you mean, so the model does not guess which ones. |
Notice the last row. A named comparison beats an adjective, because a publication is a much smaller group to average than a mood. It still helps to say which parts of it you mean.
What actually carries a voice?
Six things, in rough order of how much they change whether a reader recognises the writing as yours:
- Sentence length and its variation. The single loudest signal. AI default prose sits around 18 to 22 words with very little variance. Most human writing swings.
- Opening habits. Do you open on a claim, a question, a scene, or a number? Whatever you do, you do it most of the time, and it is the first thing a reader notices.
- Your banned list. The words you would never write. This does negative work nothing else can do, because the words you avoid are usually the model's highest probability choices.
- Structure habits. Lists or paragraphs. Subheads or none. Whether you use parentheses. Whether you ever write a one line paragraph.
- Vocabulary ceiling. The most complex word you actually use. Models drift upward from yours unless told where the ceiling sits.
- What you refuse to do. Hedge, moralise, summarise at the end, tell the reader what they are about to read. Most voices are as defined by their omissions as by their moves.
Every one of those can be written as a rule with a number or a list in it. That is the whole trick. A rule with a number survives a long generation. An impression does not.
How do I extract my style rules from my own writing?
Paste this into a fresh chat along with your samples. It asks for rules, not a description, which is the difference between something you can reuse and a compliment.
Below are [NUMBER] samples of my writing. Derive a style sheet I can reuse as instructions for you in future chats. Do not describe my writing in adjectives. I want rules another writer could follow without having read the samples. Every rule must be checkable: a number, a list, or a yes or no. Produce exactly these sections: 1. SENTENCE METRICS Average sentence length in words, the longest and shortest you found, and whether the variation is high or low. State the actual counts. 2. OPENINGS How I begin a piece and how I begin a section, stated as a pattern with the count out of the samples (for example "opens on a claim, 4 of 5"). 3. VOCABULARY Ten words or constructions I use more than most people. Five that I never use but that a writer on my topics normally would. 4. STRUCTURE Paragraph length in sentences. Do I use bulleted lists, subheads, parentheses, one line paragraphs, questions as headings. Answer each yes or no, with the count. 5. BANNED LIST Twelve words or phrases that would immediately read as not me. Include the AI defaults I clearly avoid. 6. WHAT I DO NOT DO Five moves I never make, such as summarising at the end or hedging a recommendation. 7. THE SHEET Rewrite sections 1 to 6 as 10 to 14 numbered instructions addressed to you, under 200 words total, so I can paste it into custom instructions. Flag anything you are inferring from a single sample rather than seeing across all of them. Do not flatter the writing. SAMPLES: [PASTE SAMPLE 1] === [PASTE SAMPLE 2] === [PASTE SAMPLE 3]
Two notes on the samples. Choose them for variety, not for quality: one explaining something, one persuading, one informal such as an email you actually sent. If all three are on the same subject, the model will encode the subject as your voice. Two hundred words each is enough, and more than about eight hundred adds little.
Then do the part most people skip. Read section 7 and correct it. The model will get two or three rules wrong, usually by promoting something from one sample into a habit. Deleting a wrong rule matters more than adding a new one, because a wrong rule gets applied confidently to everything you write afterwards.
Where do I save the style sheet so I stop pasting it?
A sheet pasted into one chat dies with that chat. Every major assistant has two storage places: one global, one per project. The distinction matters, because global rules apply to work they were never meant for.
| Tool | Applies to everything | Applies to one body of work | What to put where |
|---|---|---|---|
| ChatGPT | Custom instructions, under Settings and Personalization | Project instructions, set on a project | Voice rules global. Format and length rules per project. |
| Claude | Writing styles, selectable per chat | Project custom instructions | Same split. Styles can be swapped mid conversation, so keep one per channel. |
| Gemini | Saved info | Gems, one per task | Gems carry more instruction text than saved info, so put the sheet there. |
| Any of them, no setup | Not available | A note or text file you paste at the top of a chat | Works everywhere and costs you one paste. Fine for occasional use. |
Keep the global sheet short. Eight to twelve lines of rules that are true of everything you write, and nothing else. The moment you add "always end with a call to action" because you spent that week on landing copy, every unrelated answer starts ending with a call to action.
What does a working prompt library actually do about voice?
Since the argument here is that adjectives underperform, it is fair to check whether that belief shows up in the prompts we ship or only in an article. These are counts from the source files of the PrecisionPrompts library on 27 August 2026, across 455 prompts in 15 packs.
| Measure | Count | What it says |
|---|---|---|
| Prompts in the library | 455 | The sample, across 15 packs. |
| Prompts asking you to paste in your own material | 32 | More than twice as many as ask you to name a tone. |
| Prompts with a tone, voice or brand voice slot | 14 | 3 percent. The adjective is the rare tool, not the default one. |
| Of those 14, slots that supply guidance on how to fill them | 10 | A bare [TONE] label gets a bare answer, so most carry an example. |
| Of those 14, slots anchored to a named comparison | 3 | For example "like Basecamp" or "Morning Brew meets a marketing professor". |
| Prompts carrying a banned word or phrase list | 25 | Negative rules, doing work no adjective can do. |
| Prompts setting an explicit word, character or sentence limit | 92 | 20 percent. Length is the most reliable countable rule. |
| Prompts shipping a worked example output | 455 | Every one. An example shows the shape faster than any description of it. |
The 14 against 92 comparison is the honest summary of this page. When a prompt needs a specific sound, naming a mood is used sparingly and setting a countable limit is used about seven times more often.
How do I know it is working?
Run one test rather than trusting an impression, because your own writing always sounds like you to you.
- Take a piece you wrote before any of this and cut a paragraph out of the middle.
- Give the model the surrounding paragraphs and your style sheet, and ask it to write the missing one.
- Put the two versions side by side and compare sentence lengths and openings before anything else.
If the generated paragraph is noticeably smoother and more even than yours, the sheet is missing your variation rule. If it reaches for a word you would not use, that word belongs on the banned list. Each pass adds one or two rules, and the sheet is usually finished after three.
One thing to expect: the sheet gets your rhythm and vocabulary long before it gets your judgement. Voice rules govern how a sentence sounds, not which point is worth making. That part stays yours.
Frequently asked questions
Paste 3 to 5 pieces of your own writing, ask the model to derive a style sheet of measurable rules from them (average sentence length, opening habits, words you never use, how you handle lists and questions), correct anything it got wrong, then save that style sheet in custom instructions or a project so every chat sees it. Naming an adjective such as friendly or conversational does not work, because an adjective names a category millions of pages already occupy and the model returns the average of that category.
Three to five samples of 200 words or more, chosen for variety rather than quality. One sample teaches the model that topic, not your voice. Include one piece where you were explaining something, one where you were persuading, and one informal message such as an email, so the derived rules can separate your voice from that piece's subject.
Because samples alone are context, not instruction. The model reads them and still produces its own highest probability phrasing unless you convert the samples into rules it can check itself against. Make the rules countable: an average sentence length, a maximum sentence length, a banned phrase list, whether you open with a question, whether you use bulleted lists. A rule with a number in it survives a long generation. An impression does not.
Only if you store it somewhere persistent. A style sheet pasted into a single chat is gone when that chat ends. As of August 2026 the durable places are ChatGPT custom instructions under Settings and Personalization, which apply to every new chat, and ChatGPT projects, whose instructions apply only inside that project. Claude has the same split with projects and writing styles, and Gemini stores per task instructions in Gems.
Use custom instructions for rules that are true of everything you write, and paste the rest per task. Custom instructions are length limited and apply everywhere, so filling them with rules meant for one channel degrades unrelated work. A practical split is 8 to 12 lines of always true voice rules in custom instructions, with channel specific constraints such as word counts and formats kept in the prompt.
It can imitate observable patterns such as sentence rhythm, vocabulary level and structure, and the same extract and encode method works on any text you have the right to use. Two limits are worth stating. Asking for the style of a named living author usually produces a caricature assembled from what has been written about them rather than from their actual sentences. And passing an imitation off as that person's work is a legal and ethical problem however good it is. Deriving rules from writing you own, such as your company's published posts, has neither problem.
Related reading and next steps: if the output is generic as well as impersonal, the cause is usually a missing constraint rather than a missing voice, covered in why ChatGPT gives generic answers. For how long that prompt should end up being, how long should a ChatGPT prompt be measures the whole library. A job title names a category the same way an adjective does, which is measured in does telling ChatGPT to act as an expert actually work. To start from prompts that already carry the output rules, browse the prompt packs, or the Content Creator pack that several examples above come from. Banning a phrase is the smallest version of the same problem, and how to tell ChatGPT what not to do counts how our own prompts write those bans. Already bought one? The how to use guide covers filling in the bracketed variables.