How to train AI to think like you (voice, style, and reasoning, 2026)
Train ChatGPT, Claude, or Gemini to think and write like you: capture your reasoning from transcripts and samples, build a reusable voice document, and load it into custom instructions and memory.
Out of the box, an AI writes in a competent, forgettable average — the "professional but friendly" default that reads the same for everyone. Training it to think like you means giving it two things it can't infer: how you actually sound, and how you actually reason. The first is style — your sentence rhythm, word choices, and the phrases you reach for. The second is deeper and more valuable — the decision logic and mental models behind what you say, so the AI reaches your conclusions, not just your cadence.
The method below builds both into a single portable document, then loads it where the model will apply it automatically. It works across ChatGPT, Claude, and Gemini, because the artifact is text you own, not a setting locked to one platform. Two honest limits up front: an AI trained on you accelerates a first draft that still reads like you — it does not replace your judgment on what to say, and it drifts on topics you never gave it examples for. Treat the output as a draft in your voice, not a finished thought.
The steps
Separate the two goals: voice and reasoning. Decide what you are training. Voice is how you sound — cadence, vocabulary, the openings and transitions you favor. Reasoning is how you decide — the mental models, priorities, and if-this-then-that logic behind your takes. Most people only train voice and wonder why the output is on-brand but shallow. You want both, and they come from different source material, so keep them as separate tracks from the start.
Collect 6–10 real writing samples for voice. Pull together pieces you actually wrote and that were not edited by AI or a co-author — emails, posts, newsletters, memos. Six to ten spanning your normal range is enough for a model to detect your sentence length, punctuation habits, and go-to phrasings. Do not include a colleague's polished draft or a heavily-edited article; a diluted sample teaches a diluted voice.
Collect transcripts for reasoning, not just writing. Your writing shows your voice; your unscripted talk shows your thinking. Gather 3–5 transcripts from genuinely different situations — a client call, a coaching or sales conversation, a brainstorm, or a voice memo recorded while walking through a problem. Meeting tools (Zoom, Google Meet, Fathom, Granola) and voice-to-text apps capture these; wearable recorders handle in-person talk. Prefer unscripted material, where your real decision-making surfaces, over a rehearsed presentation.
Label each source with a one-line context note. At the top of every sample and transcript, add a short tag — "cold sales call", "client coaching session", "internal strategy memo". The model uses context to tell your registers apart: how you write to a prospect differs from how you talk to your team. Unlabeled files get flattened into one average voice, which is exactly the blandness you are trying to escape.
Extract your reasoning across four layers. Upload the labeled transcripts and ask the AI to pull out four kinds of knowledge: declarative (what you know and do), procedural (the step-by-step processes you follow), conditional (the decision logic for when and why you choose one path over another — your "if X, then Y" rules), and metacognitive (the mental models and how you think about the problem). Tell it to build cumulatively across files, quote specific instances, and flag assumptions you never state out loud. That last instruction is what surfaces the implicit expertise you can't easily write down yourself.
Extract a measurable style block from your writing. Now do the voice track. Feed the model your writing samples and ask it to return concrete, checkable rules, not adjectives: average sentence length, how often you use questions or fragments, contractions yes or no, em-dash vs comma habits, words and openings you overuse, and phrases you never use. "Punchy and warm" describes 50,000 styles; "sentences average 14 words, opens with a question about a third of the time, never uses 'delve' or 'unlock'" is a rulebook the model can actually follow.
Compile one portable voice-and-reasoning document. Merge the four reasoning layers and the style block into a single file — think of it as your cognitive fingerprint. A first pass may run long; condense it to the essentials that change the output, usually a page or two. Keep it as plain text you own so it survives model updates and moves cleanly between ChatGPT, Claude, and Gemini. This document, not any one platform's settings, is the asset.
Load it into custom instructions and memory. Paste the condensed document into ChatGPT's custom instructions (or Claude/Gemini's system prompt or project instructions) so it applies to every new conversation without re-pasting. Where the tool offers memory or a Project workspace, use it to hold task-specific context on top of the base voice. Layer intentionally: custom instructions set your default voice and reasoning; a Project or Custom GPT overrides it for a specific job.
Test on a known task and refine the rules. Give it a prompt you have already answered in real life and compare. Where it drifts, do not just say "more like me" — find the specific rule that is missing or wrong and edit the document. Two or three refinement rounds close most of the gap. Re-run the same test after a model update, since a new model can interpret the same instructions differently.
Common gotchas
Training only voice, never reasoning, gets you output that sounds like you but decides like the default model — on-brand and shallow. The transcript-and-decision-logic track is what makes it think like you, not just talk like you.
Vague style words ("professional, engaging, authentic") describe thousands of voices; the model splits the difference and lands on the generic default. Give it measurable rules extracted from real samples instead.
Feeding samples that were AI-edited or co-written trains a blurred voice. Use only pieces you actually wrote, unedited.
The document drifts on topics you never gave examples for — it will confidently write in your voice about things you have no stated position on. Treat those outputs as unverified drafts.
Dumping a 20-page extract straight into custom instructions wastes the model's attention and often hits length limits. Condense to the rules that actually change the output.
Setting it once and never revisiting means a model update can quietly change how your instructions are read. Re-test on a known task after major model releases.
A voice document that lives only inside one platform's settings is lost when you switch tools. Keep the source file as portable text you own.
Where Kompozy fits
The document you just built is the hard-won part — a rulebook for how you sound and how you decide. The friction is that it only helps in the one chat window where you pasted it: you still open a session, load the file, prompt for each piece, copy the result out, reformat it per platform, and post it by hand. Kompozy is where that voice document stops being something you carry into every chat and becomes the standing spec the whole engine writes to. Its Persona Brief is exactly your compiled voice-and-reasoning file, plus a banned-word filter for the phrases your style block forbids — set once, and every generation obeys it. The difference from a chat window is breadth and reach: the same brief drives 18 output formats, so one on-voice idea can come out as a text post, a carousel, a quote graphic, a blog, a newsletter, or a Persona Short of your avatar delivering it — each one already in your voice, not re-prompted one at a time. From there Kompozy fans the approved output across the 9 social platforms plus Mailchimp and blog, on a schedule, with a per-post review step so you approve or edit before anything ships — the human check this guide insists on, kept in the loop. Honest framing: if your need is one polished draft now and then, trained custom instructions in ChatGPT are enough and free. Kompozy is for when "write like me" has to become "publish like me, everywhere, every week." Creator ($49/mo, 2,500 credits) runs a solo creator's voice across a multi-platform cadence; Pro ($299/mo, 18,000 credits) covers higher volume with autopilot keeping the queue full; Enterprise is custom for agencies maintaining a distinct trained voice per client.
Frequently asked questions
How many writing samples do I need to train AI on my voice?
Six to ten pieces you personally wrote and did not have AI or someone else edit. That range is enough for a model to detect your sentence rhythm, vocabulary, and phrasing. Quality matters more than quantity — a few genuine samples beat twenty diluted or co-authored ones.
What is the difference between training voice and training reasoning?
Voice is how you sound — cadence, word choice, sentence length — and it comes from your writing samples. Reasoning is how you decide — the mental models and if-this-then-that logic behind your takes — and it comes from unscripted transcripts of you working through problems. Training only voice gets output that sounds like you but thinks like the default model.
Do custom instructions or memory work better for this?
Use both, for different jobs. Custom instructions (or a system prompt) hold your permanent voice-and-reasoning document and apply to every conversation. Memory and Projects add task- or topic-specific context on top. Custom instructions are the base layer; memory and Projects extend or override it for a specific piece of work.
Will this work across ChatGPT, Claude, and Gemini?
Yes, because the real asset is a plain-text document you own, not a platform setting. Paste it into ChatGPT's custom instructions, Claude's system or project instructions, or Gemini's equivalent. Each model may interpret it slightly differently, so a quick re-test when you switch tools is worth it.
Can AI trained on my voice just publish for me?
It can draft in your voice at speed, but it should not publish unreviewed. It drifts on topics you never gave examples for and cannot judge what is worth saying on a given day. Keep a human approval step — the training makes the first draft yours, not the final call.