AI and LinkedIn ·
How to train your AI to sound like you (without stopping being you)
"Train the AI to sound like you" is a phrase repeated so often across 2025 and 2026 that it sometimes looks like a slogan with nothing inside it. In practice it is a concrete workflow with three pieces: a well chosen corpus of your own, an instruction that captures your editorial judgement and a review habit that does not hand the last word to the model.
This guide gets into the detail of those three pieces. It is not a closed recipe, because every voice is different: it is a method you can adapt in an afternoon and that leaves you a reproducible workflow for publishing on LinkedIn with AI without it showing and without stopping being you.
By Sheena de PunkVoice · Edited by Mario Pérez
What "training" AI means in 2026
In 2023 "training" meant adjusting the model's weights with fine-tuning: technically powerful and practically out of reach for 99% of people. In 2026 the conversation has changed. Commercial models are so large and so well trained that fine-tuning adds little marginal value for most use cases; the lever that actually moves the needle is personalisation at conversation time: the context, instructions and examples you give the model every time you ask it for something.
That personalisation is called RAG (retrieval augmented generation), custom instructions, system prompt or simply "context", depending on the tool. The concept is the same: when the model writes, it does not do so from scratch, it leans on a corpus of yours that brings it closer to how you say things.
What you still need is for that corpus to be good, the instruction to be clear and your final review to impose judgement. Without all three, the model stays in the feed's average voice however many hours you throw at it.
Step 1: build your personal corpus
The corpus is the collection of your own texts the AI will be able to see when it writes in your name. The smaller the better: it does not have to be exhaustive, it has to be representative. Fifteen or twenty well chosen pieces get better results than two hundred mediocre ones.
The best materials for the corpus, in order of observed performance:
- Five to ten of your posts that did work and that you feel sound like you (not the most viral ones, the ones that most represent you).
- Two or three long personal texts: a published article, a chapter of your newsletter, a long answer in a forum or comment thread where you let yourself go.
- Voice notes transcribed literally. This is where your spoken cadence lives, which is usually more specific than your written one.
- Emails you wrote to someone you trust (with personal data anonymised). Emails capture your tone when you are not performing.
- Personal notes from a notebook: half formed thoughts, lists, drafts. They are the DNA of your mental associations.
Step 2: write the instructions that capture your judgement
The corpus gives the model your raw material. The instructions tell it what to do with it. This is where most people fail, because they write generic instructions ("write professional posts about X") instead of specific, opinionated ones.
A good instruction for your writing assistant in 2026 answers these six questions concretely. Write them in the first person and in the tone you would use with an editorial intern who has judgement of their own.
- Who you are professionally and what you stand for (three sentences at most, including two arguable positions).
- Who your reader is, not in the abstract but with the names of real examples.
- Which formats and lengths you usually use, with one example per format.
- What you never do (banned formulas, topics you avoid, tells that make you cringe).
- How you open the posts that do work (a reference to the eight hook patterns, if that applies).
- How you close the posts that do work (nothing generically motivational, always concrete).
Step 3: the review flow that keeps it from sounding like AI
With the corpus and instructions in place, the first draft already comes out respectable. But "respectable" is not "publishable". The difference comes from a short review flow you can run in five minutes that makes sure the last word is still yours.
The flow has three passes. Read it out loud first: if there is a sentence you would never say while speaking, mark it. Replace a whole section with material of your own: a scene, an anecdote, a figure only you know. Rewrite the hook and the close from scratch, without looking at the model's version, leaning only on the section you just rewrote yourself.
This flow takes ten minutes and usually removes 20% to 40% of the model's original text. That percentage is the part that will carry your real signature when the post reaches the feed.
What NOT to do when training your AI
These are the mistakes that fastest turn a promising workflow into flat output indistinguishable from anyone else's:
- Putting only other people's viral posts in the corpus "so it learns". The model learns someone else's voice, not yours.
- Writing neutral instructions ("be professional, useful and clear"). Every model is already that by default; what you add are the opinions.
- Accepting the first draft because "it already sounds fairly like me". Fairly is the trap: it is 80% your voice mixed with 20% of AI's average accent. That 20% is what smells.
- Publishing without reading out loud. It is the cheapest filter and the one that catches fastest the tells your eyes no longer see.
- Retraining every week. The corpus gets updated when you publish something that represents you better than what was there, not on a calendar.
Why PunkVoice was built on exactly this method
PunkVoice is not a thin layer over ChatGPT or a more elaborate prompt. It is the implementation of this workflow: when you sign up, it guides you to build your corpus with the pieces that do represent you, helps you write your editorial instructions in the first person and applies the three pass review flow before letting you hit "copy".
The goal is not for you to publish more. It is that every time you publish, what reaches the feed passes the platform's three filters (C2PA, statistical markers, the stylistic classifier) and, more importantly, that whoever reads you recognises your voice. Everything else (impressions, comments, followers) is a consequence of that.
Training your AI is training your own judgement
The better you curate your corpus and the more specific your instructions, the more you learn about how you write when you write well. The exercise of identifying your fifteen best pieces and naming what you do and do not do is, underneath, an editorial exercise you would have needed to do even if AI did not exist.
The added advantage is that once you have it, the AI workflow stops sounding like AI and starts sounding like you at full speed. And that combination (your own voice plus speed) is what sustains a consistent editorial calendar without it consuming your life.