Make ChatGPT better at being you

By Jack Stovell X · Instagram · LinkedIn · YouTube · TikTok · published 2026-08-19 · updated 2026-11-17

Paste two or three real samples of your writing into the chat, then describe your habits in numbers, not adjectives. Correct the draft each time it slides back to ChatGPT's average. Keep a measured prompt in your custom instructions for the baseline. Expect to repeat the samples in a new chat.

The free prompt builder below does the measuring for you. Paste 120+ words you wrote and you get a prompt in measured numbers, written for ChatGPT's custom instructions. No signup, and nothing is stored.

What custom instructions do, and where they stop

Custom instructions are free text that applies to all your chats. At the time of writing, OpenAI allows up to 1,500 characters on the Free and Go plans and 5,000 on paid plans. You find them under Settings, then Custom instructions, and the box that matters is "How would you like ChatGPT to respond?".

So you have already done the sensible thing. You wrote "conversational, direct, a bit dry, short sentences", saved it, and the first reply looked promising. Then you asked for something long and by the fourth paragraph it was back to tidy, balanced, faintly corporate prose.

That slide has a cause. A description is you interpreting yourself, then the model interpreting your interpretation. Two lossy steps. "Direct" means one thing to you and something blander to the model, and you cannot describe the habits you never notice. Under a long answer, the model's defaults pull harder than a few lines of adjectives can hold.

And to be fair, the instructions aren't wasted. They set a floor. They just can't hold a voice through a long answer, because a description holds only what you can put into words.

The how-to, inside ChatGPT itself

You can get a long way with no tool at all. This is the by-hand version, and it works for everyday drafting.

  1. Paste two or three real samples into the chat. Things you wrote yourself, with no editing pass from anyone else. Say plainly: "These are my writing. Match them."
  2. Name your habits in numbers. Count your average sentence length. Count how often you use contractions; our page on the contraction rate attribute explains how it's measured. Say "about 15 words a sentence, a contraction in most sentences, no exclamation marks" and the model has something to hit.
  3. Ask for a draft, then check it against your numbers. Count a paragraph. If the sentences are all 22 words, say so.
  4. Correct it each time it slides. Short and specific: "Too even. Mix in fragments. Cut the closing summary." Corrections work because they're concrete.
  5. Expect to redo this in a new chat. The samples and the corrections live in that conversation. Open a fresh one and you start again, so keep the samples and your numbers in a note you can paste in.

It's tedious. It does work. The weak point is step 3: you're the only scorer, and you're scoring by ear.

A measured prompt for your custom instructions

The builder above produces the same thing you'd build by hand in steps 1 and 2, only faster and from counted measurements. It takes your 120+ words and turns them into a system prompt written in numbers.

Three kinds of content go in. First, counted measurements: average sentence length, contractions per 1,000 words, commas per sentence, and the I, we and you pronoun mix. Second, judged register: formality on a scale of 0 to 10, confidence versus hedging, and your opening style. Third, hard rules that are the same for everyone, such as no dashes and no stock transitions.

A switch writes it for ChatGPT, Claude, Gemini or any assistant, so the same measurements travel with you.

Where it goes: copy the prompt, open Settings, then Custom instructions, and paste it into "How would you like ChatGPT to respond?". Check it against the character limit for your plan, then test it. Ask for a long answer, a real one, and watch whether the last paragraph still sounds like the first.

It is the one-page version of a profile. A full profile measures 45 attributes, is read live, and scores every draft. The page version can't do the last part.

Why there is still a ceiling

Anything that stores a description of your style has the same ceiling. Custom instructions, a project's instructions, a custom GPT: each holds what you can put into words, and none of them can score a draft. Numbers help, because they're more precise than adjectives. They don't change what the box is.

Some habits you can't describe because you never choose them. You don't decide how often you reach for "but" or "so". You don't pick your clause order, whether you build up to the point or lead with it. You don't plan your commas; our comma density attribute page explains how that one number is counted. These are the habits that make a reader half-recognise you without knowing why, and they're the first to go when a model drifts.

A prompt can name the ones you've noticed. It can't name the ones you haven't.

There's a second gap, and it matters more. With no score, you can't tell how close a draft landed. "Sounds about right" is a feeling. A feeling at the end of a long edit is easy to talk yourself into, and you'll ship drafts that are only partly you without ever knowing it. Better prompts narrow that gap. Only a measurement closes it.

The other route: measure it first

The alternative is to stop describing and start measuring. Take writing that already exists, three or more pieces and about 3,000 words is what we recommend, and measure it as a profile of 45 attributes across 8 groups. It usually builds in under a minute. Average sentence length and its spread, contraction rate, comma density, formality, confidence versus hedging, opening style: all counted from what you actually wrote, not from what you'd say about it.

Tools such as Acta turn a sample into a profile you can read and edit, and that's a fair approach; our comparison of ScriptGrain versus Acta AI goes through the differences without pretending there are none.

What matters is the order of operations. Measure first. Then draft under the measurement, so the profile is the constraint. Then score every draft.

The score is a voice match, 0 to 1. On-voice drafts typically land between 0.85 and 0.95, partial matches between 0.5 and 0.8, and a different voice sits under 0.45. Each score comes with deltas: per habit, the draft's value beside the profile's. So where you had "it feels a bit off" you get, for example, "sentences run 6 words longer than yours". That's a correction you can act on.

It's a style check only. It never checks facts. Scoring any text of 20+ words against a profile is free.

Connect the profile to ChatGPT

You keep ChatGPT. The profile is read live through an MCP server (41 tools) and a REST API, so ChatGPT works from your measurements and not from a pasted summary of them.

On the web, ChatGPT can add ScriptGrain as a custom MCP server, using the URL `https://mcp.scriptgrain.com/mcp`. You sign in on ScriptGrain's consent page and that's the connection made.

Whether you can do this depends on your ChatGPT plan. OpenAI limits custom connectors by plan and workspace, and it has renamed and moved the feature more than once. So we won't give you a click path here that would be wrong by next month. Our ChatGPT setup guide has the current steps.

If connectors aren't available to you, the fallback is the builder's prompt in custom instructions, as above. That route works on any plan.

Once it's connected, the main tools are `get_profile`, `generate_content` (1 credit), `check_voice_match` (free), `polish` and `save_edit` (free). One thing to know: the assistant scores a draft only when it calls the tool, so ask it to. It won't do it unprompted.

A drafting session, step by step

Here's one session for a LinkedIn post, start to finish.

  1. Name the profile. Tell ChatGPT which voice to use: "Use my profile, the one called Blog voice." It reads it with `get_profile`.
  2. Ask for the draft. ChatGPT either calls `generate_content`, which costs 1 credit, or drafts the post itself and then calls `check_voice_match`. Either way, ask for the score by name.
  3. Read the deltas. The match score tells you how close the draft landed; the deltas tell you where it missed. As an example, you might see "average sentence length: draft 21.4, profile 14.8". That's one line, and it tells you exactly what to fix.
  4. Polish. `polish` revises the draft toward a target match score (default 0.9) in up to three measured passes. It costs 1 credit, and it's free when the draft already passes. It changes how a thing is said, never what.
  5. Edit it yourself, then call `save_edit`. Your own changes are the best signal of where the profile is still off, and saving them is free.

Read the post aloud before you publish. The score says it matches your habits; only you know whether the claim is true.

Stop ChatGPT repeating you

A voice has a second problem that style can't fix: you repeat yourself. Anyone who writes weekly ends up re-making the same argument with fresh paragraphs on top.

The Voice Brain is the memory behind a voice. It keeps what your writing has already said: ideas, claims, stories and frameworks, each with its exact quote. It's free on every plan.

Before you write, run a novelty check. `check_novelty` scores a brief against the memory and tells you whether it repeats, builds on, or is new, and it proposes angles. If you're short of ideas, `what_next` proposes five angles to take next.

Once a voice has 10 indexed pieces, new drafts build on the memory and don't re-argue points you've already made. That means the fourth post on pricing can pick up where the third stopped, instead of opening with the same warm-up.

You can edit or delete any memory. If a claim is out of date, or you've changed your mind, take it out and it's gone.

What is free and what is paid

The free route costs nothing. The prompt builder on this page needs no signup, and scoring any text of 20+ words against a profile is free. The Free plan gives you one voice profile and one full analysis, plus the API, the MCP server and the Voice Brain, and you also get one free draft for the life of the account, no card.

Paid plans hold monthly credits. Writer is £12 a month for 40 credits and 1 profile. Operator is £29 for 75 credits and 3 profiles. Studio is £99 for 400 credits, 10 profiles and 3 seats. Agency is £299 for 1,500 credits, unlimited profiles and 10 seats.

A draft up to 5,000 words is 1 credit, taken only after delivery. Scoring stays free whichever plan you're on.

Questions

Can ChatGPT learn my writing style?

Within one conversation it imitates the samples you paste in, and a new chat does not reliably carry them over. Custom instructions carry a description of your style across chats, but a description holds only what you can put into words. A measured profile holds the rest.

What is the best prompt for making ChatGPT write like me?

One built from your own writing, expressed in numbers. State your average sentence length, contractions per 1,000 words, commas per sentence and pronoun mix, then add your formality and opening style. Numbers are harder to misread than adjectives. Paste two or three real samples alongside, and correct each draft as it drifts.

How many samples does it need?

For the free prompt builder, 120+ words of your own writing is enough. For a full profile we recommend three or more pieces and about 3,000 words. More text gives steadier measurements, because a single short post can be an unrepresentative day. Use writing that sounds like you at your normal pace.

Is a custom GPT or a project any better than custom instructions?

Not on the thing that limits you. A project's instructions and a custom GPT both store a description of your style, so they share the same ceiling: they hold what you can put into words, and they can't score a draft. They're useful for organising work and keeping files nearby, but they don't change the problem.

Do I need a paid ChatGPT plan?

No for the prompt route. The builder's prompt works in custom instructions on any plan, though the Free and Go plans allow 1,500 characters against 5,000 on paid plans. The connector is different: it depends on your ChatGPT plan, and OpenAI changes it often, so check the current steps before you rely on it.

Does ChatGPT send my chats to ScriptGrain?

Only what a tool call contains. When ChatGPT calls a tool such as check_voice_match, it sends the draft text that call needs, and nothing else from your conversation reaches us. If you don't ask it to call a tool, nothing is sent. You sign in on our consent page when you connect.

Does my writing train a model?

No. We never train AI models on your writing, and our terms with the AI providers forbid them from training on it. Your samples are used to measure your profile and to constrain your drafts. You can edit or delete anything in your Voice Brain whenever you like.

Related