# Your voice isn't a setting ChatGPT can turn on
The goal is right. The tooling is the problem.
Here's what custom instructions actually do
You type in a description of how you write. ChatGPT stores about five things: a tone preference, a response length preference, a communication style, a preferred language, and a free-text box for whatever you couldn't fit in the first four.
That's it. Five self-reported preferences, sitting in a settings panel.
And here's the thing: a description of your voice is not a measurement of it. Telling the model "I write casually, with short sentences" is you interpreting yourself, then asking a language model to interpret your interpretation. Two lossy translations before a single word gets generated.
So the output drifts. Under any real pressure (a long answer, a technical explanation, anything that takes concentration) the model slides back to its own average. Because that average is what it was trained to reach for. Your five bullet points in a settings panel don't stand a chance against millions of training examples pulling the other way.
The honest how-to, inside ChatGPT itself
You can get closer. Not perfect, closer. A few things genuinely help:
- Paste real samples, every chat. Two or three paragraphs of your own published writing, dropped into the conversation before you ask for anything. Not a summary of your style: the actual words.
- Name habits, not adjectives. "Punchy" means nothing to a model. Your contraction rate does. How hard your sentence length swings from one line to the next does. Say that instead.
- Correct it every single time it slides back. Not once. Every time. You're not training the model, you're babysitting it for the length of one conversation, and that supervision resets the second you open a new chat.
Do all three and the output gets noticeably closer than anything the custom instructions box produces on its own.
Why there's still a ceiling
Except there's a limit, and it's a hard one.
The markers that actually make your voice yours, function-word frequency, punctuation cadence, clause ordering, are unconscious. You didn't choose how often "so" bridges your sentences instead of "therefore". You didn't decide your semicolon rate. Nobody sits down and picks their comma density on purpose.
Which means you can't self-report it. You can't paste it into a text box, because you don't consciously know it. And here's the part that actually matters: with no score, you have no way of knowing how close any of this got. You're guessing. Every time.
A custom GPT has the same ceiling, by the way. It stores your description and your files more durably, but it is still working from a description, and it still can't tell you how close the output landed.
The other route: measure it first
ScriptGrain doesn't ask you to describe your voice. It measures it, from writing you've already published, before anything gets generated.
- Measure. Forty-five reference points pulled from your existing work: rhythm, clause patterns, punctuation habits, the stuff you can't self-report because you don't consciously track it.
- Generate under those constraints. Not "write like this person"; write inside these specific, measured boundaries.
- Score every draft against the measurement. So you're not guessing whether it sounds like you. You can see the number.
Three steps. No settings panel, no adjectives, no hoping the correction holds for the rest of the conversation.
Fair enough, but does it work
You don't need a testimonial for this one. Ask yourself the actual question: can you currently see a number that tells you how close ChatGPT's output is to your real writing?
If the answer's no, that's the whole pitch, right there.
See your own profile
Run your own writing through it. See the 45 points it pulls out, the ones you couldn't have listed yourself if someone asked. The free tier includes one full analysis with no card, and generating in that voice starts at £12/month.
At the end of the day, that's the whole difference: one route asks a model to guess at an adjective. The other gives it a measurement to hit. Your call which one you'd trust with your name on the byline.
Questions
Can ChatGPT learn my writing style?
Partially, and temporarily. Pasting real samples into a conversation gets the surface close for a while, but custom instructions only store around five self-reported preferences, the model drifts back to its own average under pressure, and there is no score telling you how close it got. The supervision also resets with every new chat.
What is the best prompt for making ChatGPT write like me?
Paste two or three real samples of your writing into the chat, name measurable habits rather than adjectives (your contraction rate, your sentence-length swing, how you open paragraphs), and correct the model every time it slides back. That is genuinely the best prompting can do; the ceiling is that your unconscious markers cannot be self-described.
How many writing samples does an AI need to learn my style?
For ScriptGrain, one piece is enough to start and about 3,000 words across several samples works best. The analysis measures 45 reference points from whatever you provide and shows a confidence score, so you can see whether more samples would help.
Is a custom GPT any better than custom instructions?
It stores more, but it has the same ceiling: it works from your description of your voice rather than a measurement of it, and it cannot score how close any draft landed. A measured profile does both, which is the difference that matters.