How many writing samples to train ChatGPT? Five of your own posts. Fewer than three and it guesses; more than ten and you're feeding it noise. Five gives it enough pattern to copy without drowning the signal.
The number is the easy part. What most guides skip is what to do with the five. So this page is the whole walkthrough. Six steps that train ChatGPT to write like you, in about thirty minutes, once.
What You Need Before You Start
Three things. Five LinkedIn posts you wrote yourself and would happily write again. A ChatGPT account (free works; so does Claude). Thirty to forty minutes, once. After that, the profile is reusable forever.
Why your own posts and not "good examples"? Because the model can only copy patterns it can see. Give it someone else's writing and it learns their voice. Generic input is why generic prompts fail, and LinkedIn's ranking now punishes exactly that: human-written posts are outperforming AI on LinkedIn because dwell time and real comments beat feed-standard polish.
Step 1: Collect 5 Writing Samples You Wrote
Do this: open your LinkedIn profile, go to your activity, and pull up your past posts. Wondering how many posts to train ChatGPT on your voice when some are weak? Pick the five that pass two tests: they performed (comments, saves, DMs), and they sounded like you when you re-read them. Aim for a mix, at least one story, one tactical post, one opinion.
Paste all five into one document. Strip out hashtags, links, and emojis; you're training on the words. Label them Post 1 through Post 5. Roughly 800 to 1,200 words total is how much writing to train ChatGPT well.
Check before moving on: read each post out loud. If one makes you wince, swap it. A weak sample teaches ChatGPT a voice you don't want.
Step 2: Run the Voice Analysis Prompt
Most prompts ask the model to write blind. This one makes it read first.
Do this: start a fresh chat, paste your five labelled posts, then paste this prompt below them:
Prompt
What you should get back: a six-part analysis quoting your own lines, ending in a VOICE RULES block. Something shaped like this:
VOICE RULES
- Open with a blunt one-line claim, never a question.
- Keep sentences under 16 words. Break the long ones.
- Use plain words: "reply", "ship", "hire".
- Never use: "leverage", "synergy", "circle back", "in today's world".
- Make the point with a short story, then one lesson.
- One idea per paragraph. Frequent line breaks. No emojis.
- End on a plain takeaway, not a "what do you think?" prompt.I ran the full version of this analysis (the engine prompt in the pack) on five of my own past posts on 6 July 2026. The rules came back specific to my writing, not boilerplate. That's the one prompt on this page I've personally tested.
Step 3: Correct the Rules It Got Wrong
The first pass is always 80 to 90% right, and the wrong 10% matters.
Do this: go through the VOICE RULES line by line and ask one question per rule: can I point to the sample that proves this? Delete any rule you can't trace to your own posts. Then add the one or two rules it missed, you'll know them on sight ("I always name the client's industry", "I never open with a number").
Check before moving on: every surviving rule has evidence in your five posts. Ten to twelve rules is plenty. More than fifteen means it's padding.
Step 4: Save Your Voice Rules for Reuse
A profile you re-paste every day is a profile you'll abandon by Friday.
Do this: put the corrected VOICE RULES block somewhere every chat can load it. In ChatGPT, that's Custom Instructions or a dedicated Project with the rules in its instructions; in Claude, a Project does the same job. Minimum version: keep the block in a note and paste it as the first message of any writing chat.
People search for a "ChatGPT voice clone", but what you're actually building is simpler and more controllable: a rules file the model follows. (It's also exactly what the Founder Voice LinkedIn Pack systemises: your profile installed as a reusable setup, plus twelve post-type prompts that inherit it. The free Voice Sketch builds a one-page starter profile if you want to try the idea first.)
Step 5: Brief Each Post with PACT
The rules fix how you sound. The brief fixes what the post says. Vague in, generic out.
Do this: for every new post, give ChatGPT four lines before it writes. PACT: Purpose, Audience, Context, Tone.
Prompt
Swap in your own four lines each time. The Context line does the heavy lifting, it's the part only you know.
Step 6: Edit the Last 20% Yourself
Never publish the raw draft. The profile plus PACT gets you 80% of the way; the last 20% is the human detail no prompt can add.
Do this, in order: read the draft out loud. Cut any line that repeats another. Swap every vague phrase for a concrete one ("optimise your workflow" becomes "block 30 minutes on Tuesday"). Rewrite the first two lines, that's where AI drifts most generic. Then add one detail only you could know: the client's exact words, the number you remember, the aside you'd make in a DM.
Here's the difference on one post, a labelled hypothetical about a hiring mistake. Generic draft:
Hiring is one of the biggest decisions a leader makes. Three lessons from a hire that didn't work out: culture fit matters, skills can be taught, trust your gut. What's yours?
Same topic through saved voice rules, then edited:
I hired for the impressive CV and ignored the flat interview. Six weeks later we both knew. The lesson wasn't "trust your gut." It was that I knew on day one and talked myself out of it.
A real moment, an admission, no tidy list. That's the profile plus your edit.
The Wrap-Up
Five samples is the answer, but the samples are only the fuel. The method is what turns them into posts that sound like you: collect, analyse, correct, save, brief, edit. Thirty minutes of setup, then every draft starts in your voice.
Do Step 1 today, just the collecting, it's ten minutes. If you'd rather start with a shortcut, the free Voice Sketch drafts a starter profile from three posts. And tell me: what's the one corporate word going straight onto your never-use list? I collect them.
Frequently Asked Questions
Can ChatGPT Learn to Write Like You?
Yes, if you train ChatGPT to write like you with your own samples and a saved rules block. It copies patterns it can see: sentence length, word choice, how you open. Five real posts, corrected rules, and your edit close the gap.
How Many Writing Samples Is Too Many?
More than about ten. Past that point the model averages across too much material and the rules come back mushier, not sharper. Five strong posts beat fifteen mixed ones every time.
Will LinkedIn Penalize AI-Written Posts?
LinkedIn doesn't penalise AI use itself, but its ranking suppresses generic, low-effort posts. Posts that earn dwell time and real comments do better, whoever drafted them.
How Do I Stop ChatGPT Sounding Generic?
Give it a voice profile built from your own posts, brief it with PACT, and always edit. Cut the hook question, the three-bullet list, and the "what do you think?" close, those are the tells. Add one detail only you know.
