How to Train ChatGPT to Write Like You: A Setup That Actually Sticks
You can usually spot AI writing in a sentence. The tidy three-part lists, the “it’s not just X, it’s Y” cadence, the word delve, the em-dashes doing all the heavy lifting. It reads fine and sounds like no one. If you use ChatGPT to draft emails, posts, or newsletters, that generic voice is the real problem — not accuracy, but the fact that it doesn’t sound like you.
The good news: you can fix this in about twenty minutes, and the fix sticks. Training ChatGPT to write in your voice isn’t a secret prompt — it’s a small setup using four features that already exist: custom instructions, a style sample, memory, and Projects. Here’s the whole thing, in order.
First, understand what “training” really means here
You are not retraining the underlying model — that lives on OpenAI’s servers and you can’t touch it. What you’re actually doing is steering it: giving it enough context about how you write that it stops reaching for its default house style and starts matching yours. Think of it less like teaching a student for years and more like briefing a very capable freelance writer before they start. The better the brief, the closer the first draft.
That distinction matters because it tells you where to spend effort: not on clever one-off prompts you retype every time, but on reusable context the tool remembers so you set it once and forget it.
Step 1: Set your custom instructions
This is the highest-leverage move, and most people never open it. In ChatGPT, go to your profile menu and open Personalization → Custom instructions. You’ll find a box asking how you’d like ChatGPT to respond. This text gets silently attached to every new conversation, so anything you put here becomes permanent standing orders.
Don’t waste it on vague adjectives like “be professional.” Be specific and behavioral:
- Name your tone in concrete terms: “Write like a smart friend explaining something over coffee — plain, warm, confident. Short sentences. Contractions are fine.”
- List what to avoid: “Never use the words delve, tapestry, testament, elevate, or navigate. No ‘it’s not just X, it’s Y.’ No bullet lists unless I ask. Go easy on em-dashes.”
- Set defaults: “Match my length — if I write two lines, don’t reply with ten. Skip the intro throat-clearing and the summary at the end.”
Those bans matter more than you’d think. Half of what makes writing sound “AI” is a handful of overused tics; naming them explicitly kills most of the tell in one shot.
Step 2: Show it your voice, don’t just describe it
Adjectives only go so far. The fastest way to transfer a voice is to paste real examples of your own writing and let the model pattern-match. Grab two or three things you actually wrote — an email you’re proud of, a post that sounded like you, a page of your notes — and hand them over with a clear instruction:
“Here are three samples of my writing. Study the rhythm, sentence length, vocabulary, and how I open and close. When I ask you to draft something, match this voice. First, describe my style back to me in five bullet points so I can confirm you’ve got it.”
That last line is the trick. Making it summarize your style forces it to actually analyze the samples, and it gives you a checkpoint to correct anything it misread before you rely on it. This “show, don’t tell” approach is the same principle behind why feeding an AI your own reference material works so well — it’s the everyday cousin of how retrieval-based AI tools look things up before they answer.
Step 3: Let memory do the remembering
ChatGPT’s memory feature can carry details about you across separate chats, so you’re not re-briefing it every morning. When it offers to remember something — or when you tell it “remember that I prefer British spelling and never open with ‘I hope this finds you well'” — that preference persists into future conversations.
Used well, memory is what turns a one-time setup into a lasting one. Used carelessly, it accumulates junk and contradictions, so check it occasionally under Personalization and prune anything stale. If you want the full picture of how memory works across the major assistants and how to control it, we walked through it in our guide to making AI actually remember you.
Step 4: Use a Project for anything you write regularly
If you write the same kind of thing often — a weekly newsletter, client emails, LinkedIn posts — create a Project for it. A Project is a dedicated workspace with its own custom instructions and its own uploaded files, separate from your general chats.
Inside one Project you can drop your style samples as files, write project-specific instructions (“this is my newsletter; always 400–500 words, one idea per issue, end with a question”), and every chat you start there inherits all of it. It’s the difference between briefing a writer fresh each time and having a colleague who already knows the account. This is also the cleanest way to keep a work voice and a personal voice from bleeding into each other.
Step 5: Close the loop — your edits are the real training
Here’s the part people skip. The setup gets you a good first draft; the polish comes from feedback. When a draft comes back slightly off, don’t just silently rewrite it yourself — tell the model what was wrong: “Too formal, and you used ‘moreover’ — I never say that. Loosen it up and cut the last paragraph.”
Do that a few times and two things happen: the current draft gets sharper, and — if you fold the recurring corrections back into your custom instructions — the next draft starts there instead. Treat it like editing a junior writer whose only job is to get better at sounding like you. That loop, repeated, is where the voice actually locks in.
The honest limits
A few things to keep expectations level. It will drift — over a long chat it slowly slides back toward its default voice, so for important pieces, start fresh or re-paste a sample. It’s pattern-matching, not you — it can imitate your rhythm, but it doesn’t have your memories or your actual opinions, so the ideas and the facts still have to come from you, and you still have to read every word before it goes out. And mind your privacy: don’t paste anything confidential — client secrets, personal data, unpublished work you care about — into samples or memory unless you’re comfortable with it living in your account. If any of that gives you pause, that’s a fair reason to look at running a model on your own machine instead; there’s a growing wave of small AI models you can run locally and privately.
Done right, though, this is one of the highest-return twenty minutes you’ll spend with any AI tool. You stop fighting the generic voice and start getting drafts that already sound like you on the first try — which means less rewriting, faster output, and writing that’s still recognizably yours. The tool doesn’t replace your voice. Set up well, it just gets out of the way and lets you use more of it.
Sources & further reading:
Related Reading
- Make AI Actually Remember You: A 2026 Guide to Memory in ChatGPT, Claude, and Gemini
- The Everyday AI Prompts That Actually Save You Time
- ChatGPT Scheduled Tasks: Put Your AI on Autopilot for the Work You Keep Forgetting