How to Make ChatGPT Write LinkedIn Posts in Your Voice, Not Its Own

You can spot one from the first line. "I almost didn't post this." A line break. "But here's the thing." Three short paragraphs that each end on a small drumroll, a list where every item starts with an emoji, and a closing question nobody asked for: "What would you add? 👇"

Nothing in that post is wrong, exactly. It just isn't anyone's. Give ChatGPT a topic and nothing else, and it writes the most likely LinkedIn post about that topic. Your readers have already scrolled past that post a thousand times.

I build writing tools for LinkedIn and draft with ChatGPT and Claude most days, so I've spent more time on this problem than I'd like. What finally worked was simple: stop describing my voice and show it, keep everything in one Markdown file I own, and feed the model the most useful example there is, my own edits.

Why "write it in my tone" doesn't work #

The usual fix is one line in the prompt: write in a conversational, direct, slightly witty tone. Read it again and ask who it doesn't describe. Almost everyone writes some version of it, so the model maps it to the same average post it would have written anyway, plus a joke.

Voice lives in things that are hard to describe and easy to show: how long your sentences run, which words you never use, whether you open with a scene or a claim, how you end. Models pick these up from examples far better than from adjectives, and two studies from 2025 measured how much better:

• In one, models imitating an author with no samples to go on matched the style in under 7% of attempts, as judged by an authorship classifier. Given a few samples, they did up to 23.5 times better.
• In the other, five samples beat none on every measure, while going from two samples to ten added little.

The second study also has a sobering result: on informal writing like blog posts, most of the imitations still failed the authorship check. Examples get the model most of the way. Your own edit does the rest, and nothing in this guide replaces it.

The patterns that give an AI post away #

Before you build anything, know what you're steering away from. The most thorough public list is Wikipedia's Signs of AI writing, written by editors who clean chatbot text out of articles. Most of it carries over to LinkedIn, and the feed adds a few patterns of its own:

One caution from the Wikipedia guide applies here too: people wrote like this long before chatbots did, so a single pattern proves nothing. Passing a detector isn't the point either. These patterns matter because your readers have learned to skim past them in a fraction of a second, whoever wrote them.

Build a voice file #

A voice file is one document that holds what a model needs to sound like you. Write it in Markdown. It's plain text that every model reads, it pastes into any instructions box and uploads as a file anywhere, and it stays yours when a product changes (more on why that matters below). It has five parts:

1. Who you are and who reads you. Two or three lines: your role, your readers, and what you would never claim. This stops the model from inventing a career for you.
2. Your posts. Five to eight you wrote yourself, pasted whole, first line included. Choose by comments and reposts rather than likes; your post analytics show which ones landed. Mix lengths and topics so the model learns the voice, not the subject.
3. Patterns, not adjectives. Habits anyone could check in the text: sentence length, how you open and end, punctuation you use or never use, words you lean on.
4. A never list. Your own tells: the rows of the table above that apply to you, plus whatever you keep deleting from drafts.
5. Before and after. Pairs of an AI draft and the version you posted. The next section is about these, because they do the most work.

The fastest way to write part three is to let the model read part two. Paste your posts into a fresh chat and ask:

List 12 patterns in these posts that anyone could check in the text: sentence length, how they open and end, punctuation, words used often, words never used, how paragraphs are built. No adjectives about tone.

Then delete every line that isn't true. What's left will be more precise than anything you'd write about yourself, because people rarely notice their own habits. Here is the shape of the finished file:

# Voice file: [your name]

## About me
Head of support at a 60-person B2B software company.
I write for support leads and founders. I don't manage
engineers and never claim to.

## Patterns
- Sentences run 8 to 20 words.
- I open with what happened, not with the lesson.
- "I", not "we", unless I mean my team.
- Paragraphs are one to three sentences.
- No exclamation marks, no hashtags, no emoji bullets.
- I end on the point, never on a question to the reader.
- Numbers come from my own work, rounded honestly.

## Never
- "It's not X. It's Y."
- Three adjectives in a row
- delve, landscape, navigate, unlock, game-changer
- "Here's the thing", "Let's talk about", "Thoughts?"

## My posts
### 2026-09-14
[the full post]

## Before and after
### AI draft
[the draft]
### What I posted
[your version]

Your edits are the best examples you have #

Every time you rewrite an AI draft before posting it, you create an exact record of your voice. Each cut is a rule, and each swapped word is a preference. One draft placed next to your final version teaches a model more than a paragraph of instructions, because it shows where its defaults and your taste part ways. Put two or three such pairs in the voice file and tell the model the second version of each is the target.

The catch is finding them later. The drafts sit in a chat history that grows by dozens of chats a week, and the official data exports arrive as a ZIP of JSON files. You can copy each good drafting chat into a note by hand, and that works if you're disciplined about it. I wasn't, so I built Chat to Markdown (also mine). It saves ChatGPT, Claude and Gemini chats as Markdown files in a folder you choose, such as an Obsidian vault, and updates the file when you return to a chat. The whole history of a post, from the first prompt to the draft you settled on, ends up in one file you can search.

However you keep them, the routine is the same. Once a month:

1. Open last month's drafting chats next to the posts you actually published.
2. Pick the two or three where you changed the most.
3. Add each one to the file as a before-and-after pair, and drop the oldest pair once you have more than four.
4. Move any phrase you deleted twice to the never list.

It takes about ten minutes a month, and the drafts need fewer edits each time. That is the only test of a voice file that matters.

Where the voice file goes in ChatGPT, Claude and Gemini #

All three main assistants let you attach standing instructions and files to a workspace, so you don't paste the voice file into every chat. As of November 2026:

Split it the same way everywhere. Put the short rules in the instructions box, which the model reads on every turn, and the long material, posts and pairs, in the file. Start each new post as a new chat inside the project, so the drafts of one post don't leak into the next.

Gemini shows why the master copy belongs outside any one product. Google is replacing Gems with skills: on personal accounts, Gems become read-only on November 17, 2026 and disappear from the app on December 17. Workspace business accounts follow in March 2027, education accounts in June. Google converts them automatically and lets you download a Gem's files before the cutoff, but if your voice lived only in a Gem, copy it out this week. A Markdown file on your own disk survives every rename, and switching assistants takes one upload.

Draft from material, not from a topic #

Even with a good voice file, "write a post about hiring" gets you a post about hiring in general, in your voice. Voice is only half of it. The other half is material only you have: what happened, who said what, the number that surprised you. Give the model that, however rough. A voice memo transcript or bullet notes typed on your phone will do.

The prompt I use inside the project:

Turn the notes below into a LinkedIn post that follows my voice file. Before you write, ask me up to three questions about anything vague or missing. Keep the first two lines under 140 characters and make them the point, not a warm-up. Plain text only: no markdown, no headings, no hashtags. Give me two versions with different openings.

Notes: [paste]

The questions step does the most work. A model can't make up your Tuesday, but it can notice that your notes skip the moment the client changed their mind, and ask about it. The two openings give you a choice where it matters most: a phone shows only about 140 characters before "…more", so the first two lines decide whether anyone reads the rest. The see more cutoff guide covers exactly where that line falls.

The ten-minute edit #

What comes back is a draft. These steps make it yours:

1. Delete the first sentence and read the post without it. Most of the time it was a warm-up, and the post improves.
2. Search for your never list with Ctrl+F (⌘F on a Mac). It catches more than you'd expect.
3. Read it out loud. Cut anything you wouldn't say to a colleague across the table.
4. Add one detail only you could know: a name, a figure, the exact words someone used.
5. Check the format. LinkedIn renders no markdown, so pasted asterisks stay asterisks (how to paste from ChatGPT covers the fixes). Bold two or three phrases at most. The Post Formatter extension converts formatting as you paste, and its Hook check grades the lines above "…more" before you post. The free LinkedIn post preview shows the same cut without installing anything.
6. When a post does well, add it to the voice file. Then the file keeps learning from what works with your readers.

What this won't fix #

A voice file makes drafts sound like you. It doesn't give them anything to say. If the notes are thin, the post will be a well-dressed version of thin. The research is also plain about the ceiling: even close imitations stayed statistically more predictable than human writing. That's fine for a draft, and a good reason to keep the last pass for yourself.

The bottom line #

Stop describing your voice to ChatGPT and show it: five to eight of your own posts, habits anyone could check, a list of phrases you never use, and AI drafts next to what you actually posted. Keep it all in one Markdown file you own, load it into a ChatGPT or Claude Project or a Gemini skill, give the model material instead of topics, and save your edits so the file gets better every month.