# Train AI on your writing style · three routes compared · ScriptGrain > Prompting drifts and fine-tuning is heavy, expensive and unverifiable. A measured profile takes one sample and about 45 seconds, then scores every draft. Your writing is measured, never trained on. Canonical: https://scriptgrain.com/us/train-ai-on-your-writing-style # Train the AI you already use on your writing style *By Jack Stovell · published 2026-08-21* You can prompt a model with samples. You can fine-tune one. Or you can measure your voice properly, hand the measurement to ChatGPT, Claude or whichever assistant you use, and score what comes back. Only one of these holds up past the first paragraph. ## The three routes, compared honestly **1. Prompting with samples: free, fast, shallow.** Paste in a few paragraphs and the model apes your surface for a bit. Then it drifts back to its own average. Every new conversation, you start from zero again. **2. Fine-tuning: real, but heavy.** This one actually works, in theory. You need a large pile of your own text, money, and a fair bit of technical setup. Your style gets baked into a model you now have to host and maintain. And here's the kicker: at the end of it, you still can't check how close any single draft actually got. No ruler. Just vibes. **3. A measured voice profile: light, and checkable.** This is what ScriptGrain does. Forty-five reference points pulled from as little as one sample, in about 45 seconds. That profile becomes a hard constraint when generating and a ruler afterward. Every draft gets scored against it. Not "does this feel like you", but "does this match, point by point". ## Why prompting alone doesn't hold Here's the thing about copy-pasting samples into a chat window: it's not measurement, it's mimicry. The model glances at your tone, borrows a few tics, then settles back into whatever its training data thinks good writing sounds like. Shallow, and short-lived. Fair enough for a one-off email. Useless for anything you need consistently. ## Why fine-tuning solves the wrong problem Fine-tuning is the industrial option. Heavy infrastructure for a problem that doesn't need it. You want your sentence rhythm captured, your comma habits, the way you open paragraphs; you don't need a maintained model sitting on a server somewhere. And even after all that spend, there's no scorecard. You can't point at any one draft and check the match. You're just hoping the weights learned something. ## What a voice profile actually measures Lexical choices. Sentence syntax. Tone and register. Rhetorical structure. Punctuation habits. Function-word patterns. Cadence. Forty-five points, not a vague impression. That's the difference. A profile is legible. You can look at it, argue with it, add samples and re-analyze when it reads wrong, and hand it to the generator with "write against this". A fine-tuned model is a black box that cost you a stack of cash to build. ## Your privacy, stated plainly ScriptGrain does not train a model on your writing. Full stop. Your samples get measured, not absorbed. We never train AI models on anything you upload, and the providers that run the analysis are contractually barred from training on it. The profile lives as a set of extracted, readable attributes, not as weights baked into a neural network you can't inspect. To be fair, that distinction matters more than people think. Training absorbs. Measuring extracts. One of those can be handed back to you in a readable form. The other can't. ## How much you actually need One sample works. About three thousand words spread across several samples is best, if you've got them lying around, and you can add up to twenty. That's it. No dataset wrangling, no weeks of collection. ## See it for yourself Paste in something you've written. One piece, an email, an old blog post, whatever's closest to hand. In about 45 seconds you'll get your own profile back: the lexical fingerprints, the syntax habits, the rhythm you didn't know you had. The free tier includes one full analysis with no card required; writing with the profile starts at $17/month. At the end of the day, prompting drifts, fine-tuning costs a fortune and tells you nothing afterward, and a measured profile does both jobs properly: it constrains the generation and it scores the result. That's not a close call. ## Questions ### Do you fine-tune a model on my writing? No. ScriptGrain measures your writing into a profile of 45 readable attributes; nothing is trained on it. We never train AI models on your samples, and the providers we send them to are contractually barred from it. The profile is applied as a constraint at generation time, which is also why every draft can be scored against it. ### How many writing samples do I need? One piece is enough to start. Around 3,000 words across several samples gives the best results, and a profile accepts up to twenty samples. The analysis reports a confidence score, so you can see when adding more writing would sharpen it. ### Can it learn a brand's voice rather than my personal one? Yes. A profile can be built for any voice you have writing samples for, per person or per brand, and paid plans include multiple profiles so an agency or team can keep several voices separate without cross-contamination. ### What happens when my style changes? The measured attributes are read-only by design, so a profile always reflects what was actually in your samples. When your writing moves on, add newer samples and re-analyze, or build a fresh profile from them. Saving your final edits also feeds ScriptGrain's closed-loop learning, so drafts track the voice you actually publish, not the one you started with. ## Related - [AI that writes like you](https://scriptgrain.com/us/ai-that-writes-like-you) - [Make ChatGPT sound like you](https://scriptgrain.com/us/make-chatgpt-sound-like-you) - [Free writing style analysis](https://scriptgrain.com/us/tools/writing-style-analysis) - [Why ScriptGrain](https://scriptgrain.com/us/why)