How to train ChatGPT to write in your voice

What does "training" ChatGPT actually mean?

When people talk about training ChatGPT on their voice, they usually mean getting it to write like them without being told each time. The model itself doesn't change, and nothing you type retrains it. What changes is the context it receives: the text it reads before it starts writing.

ChatGPT gives you several places to keep that context. Custom instructions apply across your chats. Projects group conversations together with their own instructions and files. Memory lets it carry certain details from one conversation to another. Custom GPTs package instructions and reference files into a reusable assistant. Exact options vary by plan and change over time, but underneath they all do one job: put your description in front of the model before it writes.

That tells you where the effort belongs. The result depends almost entirely on the quality of the description. A vague description stored carefully still produces vague writing.

Why pasting samples only gets you halfway

The obvious shortcut is to paste five of your posts and say "write like this". It does something. The output picks up your topics, some of your phrases, perhaps your paragraph length.

What it rarely picks up is why you wrote the way you did. It sees that one post opened with a customer complaint and starts opening everything with a customer complaint. It notices a phrase you used twice and works it into every piece. It copies the surface of the examples, including details that belonged to one occasion, and misses the judgement that produced them.

Samples also anchor on subject matter. If all five are about pricing, the next piece about hiring drifts towards pricing language. They also take up a lot of context for what they teach, and that space is shared with your actual request.

Samples are far more useful as raw material. Read them yourself, work out what they have in common, and write that down.

Build a voice document from your best writing

This takes an hour or two. You do it once and revise it occasionally.

1. Choose the pieces. Pick a handful of things you'd happily put your name to today: emails that got the reply you wanted, posts that drew comments for the right reasons, a proposal that won. Mix formats if you can, and leave out anything heavily rewritten by someone else.

2. Read them for decisions. For each piece, note what you actually did. How did you open? How long were most sentences? Where did you use a technical term, and where did you explain one? Did you admit uncertainty? Did you name a price, a competitor, a mistake you'd made? What did you leave out?

3. Turn qualities into behaviours. Most people skip this step, and it matters most. If your instinct is "I'm direct", ask what direct looks like on the page. Perhaps it means the recommendation goes in the first line, you don't apologise for disagreeing, and you rarely write "might". Those are instructions a model can follow. "Direct" gives it nothing to do.

A worked example. An invented freelance bookkeeper reviews her best client emails. Her first description is "friendly but professional". After reading them for decisions, she ends up with this:

• Opens with the answer to the client's question, then gives the reasoning
• Writes "you'll need to" for actions, never "you may wish to consider"
• Explains each tax term the first time it appears, in brackets, in under ten words
• One light remark per email at most, usually about the paperwork
• Ends on the exact next action and a date

That list will move ChatGPT's output far more than "friendly but professional" ever did.

4. Note the words. Write down the vocabulary you reach for and the vocabulary you avoid: industry terms you use with customers, phrases that make you wince, words your competitors wear out.

5. Describe the reader. Write a short paragraph on who usually reads you, covering what they are trying to get done, what worries them and how much they already know. If you write for very different readers, write a paragraph for each and label them.

What the finished document looks like

Keep it short enough that you would read it yourself. A page is plenty. A sensible order runs like this:

• Two lines on who you are and who you write for
• Your behaviours, written as instructions
• Words and phrases to use, and ones to avoid
• Structural habits: openings, typical length, endings
• One or two short examples, each with a line on what it demonstrates

The annotation on those examples matters. "Note that the opening states the cost before the benefit" teaches a behaviour. A bare example mostly teaches a topic.

Where to keep it so it gets used

The document only works if the model sees it every time. Pasting it at the top of each chat works, but depends on you remembering. Custom instructions apply it broadly, which suits one person writing in one voice. A project suits a defined body of work, with related files kept alongside. A custom GPT is the easiest way to share the same setup with a colleague.

Wherever it lives inside ChatGPT, keep the master copy in a document you control. Tools change their settings and limits, and your voice description shouldn't exist only inside one of them.

Test it against something you've already written

Pick a piece you wrote last month. Give ChatGPT the brief you had then, with the voice document in place, and put the two versions side by side. Mark every sentence in the AI version that you wouldn't have written. Each mark points to an instruction that is missing or too vague. Add a line to the document and run it again.

Two or three rounds usually close most of the gap. After that, revisit the document when the business shifts, when you start serving a new kind of customer, or when you notice yourself making the same correction in every draft.

Mistakes that weaken a voice document

A few problems turn up again and again.

Writing the aspiration. People describe the writer they would like to be, and the model faithfully writes like someone who isn't them. Describe what your best work already does. You can push it further later.

Too many rules. A document with forty instructions gets followed unevenly, and some rules will contradict each other in ways you won't spot until the output goes strange. If two rules point in different directions, decide which one wins and delete the other.

Only negatives. A list of banned words and phrases tells the model what to avoid and nothing about what to do. For every "never", try to write the "instead".

One reader for everything. If you write tenders for hospital procurement teams and posts for café owners, a single audience paragraph will produce writing pitched at neither. Label each audience and say in your request which one a piece is for.

Never revisiting it. Businesses change what they sell and who they sell to. A voice document written two years ago can quietly pull new writing back towards an old positioning. Put a reminder in your diary to reread it every few months.

When one document stops being enough

A single voice document works well for one writer, one audience and a handful of formats. Add more and it strains. Several audiences get squeezed into one paragraph. Channel rules for LinkedIn and email get tangled up with brand rules. Colleagues keep their own copies, and the copies drift apart.

That is where a structured profile earns its place. In BraVo the voice document becomes a profile with separate parts: the business, the brand (archetype blend, voice and tone, values), and up to three audiences, each kept in its own record so a tender reader and a café owner never share a paragraph. BraVo stores that profile and applies it to every piece; it doesn't learn from or train on your content. Channel rules live in their own layer too, so LinkedIn conventions stop leaking into your brand rules. You still edit what it writes.

If you are starting today, start with step three. Take the one adjective you'd use for your writing and write down three things it looks like on the page.