Capture a brand voice from writing you already have, and stop calling it training
A prompt that reads your best existing writing and returns a reusable style instruction block, with a quotation behind every rule it gives you.
- Works in
- ChatGPT, Claude, Gemini
- You need
- Three to five pieces of your own published writing · Who wrote them · The formats the voice has to work in
- Written for
- how to train chatgpt for seo
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
Nothing you paste into a chat window changes the model. You are supplying examples in the context, for that session, and the useful version of this job is therefore not training at all: it is producing a block of text you keep, that describes how your writing behaves, backed by quotations from writing that already exists.
Why the word matters
Pasting samples into a chat does not modify the model. It adds text to the context the model reads, and when the conversation ends nothing about the model has changed.
Custom instructions, project instructions and memory features are storage rather than learning. They hold text and replay it into the context automatically so you do not have to paste it each time, which is genuinely useful and is not the same thing as the model having learned your voice. Fine tuning does change weights, and it is a developer process with a prepared dataset behind it, not something happening when you upload three blog posts.
This is worth being blunt about because it changes what you should build. If the voice lives in your context rather than in the model, the artefact you want is portable text you own, testable against a draft, movable between tools, and readable by a new writer on their first day.
Adjectives do not constrain a sentence
A usable rule names something checkable in a draft. “Conversational and authoritative” is not that: two writers reading it produce different paragraphs, and so do two models.
The rules that survive contact with an actual draft look like constraints. Sentences under 25 words. Never open a section by restating its heading. State claims flat rather than hedging them. Concede the objection before answering it. Every one of those can be verified by reading the output, which is why rule 1 of the prompt insists on a quotation behind each, and why the rules a model invents without evidence are the ones to delete.
What a voice block cannot carry
Judgement, and having something to say. The block reproduces rhythm, person and refusals, which is the surface, and the surface was never the reason a draft felt empty.
A voice block applied to a piece with no specifics in it produces prose that sounds like you and contains nothing, and that is a worse outcome than an obviously generic draft because it survives a skim. Pair it with the drafting prompt, which refuses to write until you hand over a real example, and keep the two jobs separate: one decides what goes in, the other decides how it sounds.
You are analysing writing samples in order to produce a style instruction I
can paste at the top of future sessions. Derive everything from the samples in
front of you. Do not apply general advice about what good writing looks like.
Work out only what the evidence supports:
- Sentence length, and how much it varies within a paragraph.
- Whether the writing uses first person, second person or neither, and how
consistently it holds that.
- Contractions: used, avoided, or mixed.
- How a section opens. Quote three opening sentences.
- What the writing does with a claim: hedges it, sources it, or states it flat.
- Constructions that recur across more than one sample.
- What this writing never does. This section is the most useful one and it is
the one you will be tempted to leave empty.
Rules you must follow:
1. Every rule you state must have a quotation from a sample sitting next to
it. A rule you cannot evidence does not go in, however true it sounds.
2. Do not describe the voice with adjectives. "Conversational",
"authoritative", "friendly" and "engaging" are useless to a model because
they do not constrain a sentence. State rules that could be checked against
a draft by someone who has never read the samples.
3. If the samples disagree with each other, name the split and say which
samples fall on each side. Do not average them into a voice that appears in
none of them.
4. If I have given you fewer than three samples, or they are all the same
format, say what you cannot conclude from what you have and name the piece
of writing that would settle it. Do not fill the gap with a guess.
5. Do not invent anything about my company, my products, my customers or my
opinions. You are describing how the writing sounds, not what it says, and
nothing about the subject matter belongs in the output.
Output in this order:
1. A fenced code block headed STYLE INSTRUCTION. Inside it, 12 to 20 numbered
rules written as second person instructions to a model, with no adjectives
and nothing you could not evidence. This block has to work on its own in a
session that has never seen the samples, so it must not refer to them.
2. The evidence. Each rule number, then the quotation it came from.
3. What you could not determine from these samples, and which piece of my
writing would settle each one.
My samples: [PASTE THREE TO FIVE PIECES YOU WOULD BE HAPPY TO SEE COPIED, SEPARATED BY A LINE OF THREE HYPHENS]
Who wrote them: [ONE PERSON, A TEAM, OR MIXED, AND WHICH IS WHICH]
Where this voice has to work: [THE FORMATS YOU WILL USE IT FOR]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[PASTE THREE TO FIVE PIECES YOU WOULD BE HAPPY TO SEE COPIED, SEPARATED BY A LINE OF THREE HYPHENS]- Your best writing, not your most recent. Three to five pieces of roughly 800 words each is enough, and one of them should be something short and awkward, like a support reply or a refund email, because that is where a voice shows what it does under pressure. Samples written by an agency are not your voice, they are theirs.
[ONE PERSON, A TEAM, OR MIXED, AND WHICH IS WHICH]- Decides what rule 3 does. If you write "mixed, samples 1 and 2 by the founder, 3 to 5 by the content team", the model reports the split instead of blending two voices into an average that sounds like neither. Blending is how house style ends up describing nobody.
[THE FORMATS YOU WILL USE IT FOR]- Say "long form articles and product page copy" or "LinkedIn posts and sales emails". A voice that works in a 1,500 word article does not survive a 40 word push notification, and stating the format stops the model deriving rules about section openings that will never apply to what you actually write.
How to run it
- 01Pick samples you would be happy to see copied
Choose three to five pieces you are genuinely proud of rather than the last five things published. Whatever is in these samples is what comes back multiplied, so a mediocre sample teaches the model your mediocre habits with the same fidelity as your good ones.
- 02Strip everything that is not the writing
Remove navigation, bylines, calls to action, boilerplate footers and anything the CMS added. The model cannot tell a template from a sentence you chose, and leaving the boilerplate in produces a rule about how your writing always ends with a demo booking link.
- 03Run the prompt and read the evidence column first
Go to the numbered evidence before the style block. Any rule whose quotation does not obviously support it is a rule the model reasoned its way to rather than observed, and it will quietly distort every draft you generate from that block. Delete those rules by hand.
- 04Save the block somewhere you own
Put the STYLE INSTRUCTION block in a document, a snippet manager or your repository, not only in a chat product setting. It is portable text: the same block works in ChatGPT, Claude and Gemini, and it survives you changing tools, which an account level setting does not.
- 05Test it against a piece it has never seen
Take an article you wrote after the samples, hide it, and ask a model with only the style block to write on the same question. Compare the two. The gap you see is exactly what the block failed to capture, and it is usually the part where you concede something.
- 06Check that the voice did not cost you the structure
Distinctive voices tend towards long unbroken paragraphs and openings that build. Run a published page through the AI content readiness check to confirm the answer still arrives in the first sentence under each heading, because that is what decides whether an assistant can lift a passage out.
Questions people ask
Can you actually train ChatGPT on your brand voice?
Not in the sense the word implies. Pasting samples into a chat, saving custom instructions or adding files to a project all put text into the context the model reads, and none of them change the model itself. Fine tuning, which does change weights, is a separate developer process with a prepared dataset and is not what happens in a chat window.
How many writing samples are enough?
Three to five pieces, covering at least two different formats. Below three the model cannot tell a habit from a one off, and above five or six the rules stop changing while the context you have to carry keeps growing. Range matters more than quantity: one short reply teaches more than a fourth long article.
Why should the style instruction avoid adjectives?
Because an adjective cannot be checked against a sentence. "Authoritative" produces different output from every model and every writer who reads it, so it constrains nothing. "State claims flat, without hedging verbs" and "never open a section by restating its heading" are rules a draft either follows or does not.
Where do I put the style instruction once I have it?
In a file you control, and then into whatever context mechanism your tool offers: custom instructions, project instructions, a system prompt, or the top of each conversation. Keeping the file as the source of truth means the voice moves with you rather than living inside one account setting you cannot export.
Will a voice block make AI writing sound human?
It fixes the surface and not the substance. A model given your voice and no facts produces fluent generality that now sounds like you, which is arguably worse, because it passes a skim. The specifics that separate written from generated still have to come from the writer.
Unsubscribe in one click. We never pass your address on.