Can ChatGPT or Claude write LinkedIn posts in your voice?
Yes, for first drafts, if you give it three things: posts you wrote yourself, written voice rules, and the raw material for each post. Without them, both tools fall back on the same polished, interchangeable style, the kind of generic content LinkedIn said in May 2026 it would dial back.
The announcement, published by Laura Lorenzetti on 20 May and republished in LinkedIn's Pressroom on 4 June, says content that appears AI-generated and lacks clear perspective "is less likely to be widely distributed beyond a person's immediate network". Our post on what LinkedIn's AI-content update means for reach covers the rest of that announcement.
LinkedIn's help page on content created with AI draws the line clearly. AI is fine for refining language, proofreading or being more concise, but the post should reflect your own voice, perspective and experience. LinkedIn urges you to review, edit and approve anything AI helped with, and recommends telling readers when you relied heavily on AI, unless that is obvious from the context. Members can also flag a post as "Seems like AI slop", and if a post gets enough of that feedback you see a tip in its analytics.
Where should your voice live in ChatGPT and Claude?
Put it in a project. Both tools have account-wide instructions, project instructions and project files, and a project keeps your LinkedIn rules away from everything else you use the assistant for. Here is what each help centre says, checked 24 October 2026:
| Setting | ChatGPT | Claude |
|---|---|---|
| Account-wide instructions | Custom instructions apply to all chats. Free and Go: up to 1,500 characters. Plus, Pro, Business, Enterprise and Education: up to 5,000. | Instructions under Settings apply to all your conversations. |
| Project instructions | Apply only inside the project and override your global custom instructions. | Apply only to chats inside that project. One Claude help page also lists project instructions as paid plans only, though it says free accounts can create projects. If you can't set them, upload your voice rules to project knowledge. |
| Project files | Upload PDFs, docs or pasted text. Free plans allow 5 files per project, Go and Plus 25, and Pro, Business, Enterprise and Edu 40, so put your sample posts in one file. | Anything in project knowledge is used across all chats in that project. Free accounts can create up to five projects. |
| Keeping good drafts | Add to project sources saves a response for reuse in later project chats. | Context is not shared across chats in a project unless you add it to project knowledge. Each project also has its own memory. |
Two rows matter more than they look. ChatGPT project instructions override your account-wide ones, so a general setting such as short answers in bullet points won't leak into your posts. And Claude's note about context means a correction you type in Monday's chat is not guaranteed to reach Thursday's chat. Project memory may pick some of it up, but a saved file of corrections is explicit. The correction log later in this guide deals with that.
Which writing samples should you give it?
Start with 5 to 10 posts you wrote without help, chosen because they sound like you. How well they did matters less. Advice on the number varies:
- 3 to 5 examples. Anthropic's prompting best practices for Claude recommend 3 to 5 examples for best results, ones that are relevant, diverse and wrapped in <example> tags so Claude can tell them apart from instructions.
- 5 to 10 posts. ContentIn, which sells an AI ghostwriting tool, suggests five to ten of your strongest LinkedIn posts in one text file, picked by recognition: posts readers said sounded like you. It warns against content you didn't write.
- 25 posts. Brendan McNulty, who runs the 52 AI experiments newsletter, gave Claude 25 earlier LinkedIn posts plus a podcast transcript in an experiment published on 28 May 2025.
In our view, the counts differ because the jobs differ. Three to five examples set a format. Ten or more show the range: how you open a story post compared with an opinion post, and how you end each one. Cover each format you use at least twice so the AI doesn't copy one post's quirks into everything.
Leave out posts a ghostwriter or an earlier AI tool wrote, posts that are mostly a link, and anything you'd no longer say. If you have fewer than five posts, use emails you sent to customers or a transcript of you explaining your product. In our view, spoken words are closer to your voice than a polished About section.
How do you write voice rules the AI will follow?
Write short, specific rules and give the reason for each. Anthropic's guide makes two points that apply to any assistant. First, explain why a rule exists. Its example swaps a bare ban on ellipses for a rule that says the text will be read aloud by a text-to-speech engine, and notes that Claude can generalise from the explanation. Second, tell the model what to do instead of what not to do. ContentIn takes the opposite view and says don'ts matter more, so keep any ban you need and pair it with the word or habit to use instead.
Copy this into your project instructions and fill in the brackets:
| Voice rules template (about 900 characters before you fill it in) |
|---|
| Who I am: [Name], [role] at [company]. I write for [buyer role] at [type of company]. |
| Topics: [topic 1], [topic 2], [topic 3]. Nothing outside these. |
| Tone: [one word, for example blunt]. I sound like a peer explaining something over coffee. |
| Sentences: mostly under 15 words. Paragraphs of 1 to 3 lines. |
| Openings: start with a number, a mistake I made, or a claim some readers will disagree with. |
| Endings: one question readers can answer from their own work, or no question. |
| Phrases I really use: [3 to 5 phrases]. |
| Plain words: write "use", "help", "start" instead of longer business words, because my readers skim on their phones. |
| Facts: use only the facts, numbers and stories in my notes. If something is missing, write [ASK ME] in its place. |
| Format: plain text, no hashtags, no emoji, under 1,300 characters. |
| Check: compare the draft with my example posts and cut any line I would not say out loud. |
The [ASK ME] line matters most. It gives the AI a way to flag a gap instead of filling it with an invented client, number or story, which is the fastest way to publish something untrue under your name. For opening lines, our collection of LinkedIn hook examples gives patterns to name in the "Openings" rule.
What prompt turns your notes into a post?
Give the AI your raw material and ask for one post. Your notes go first, then the examples, then the request, which is the order Anthropic's guide recommends for long inputs of 20,000 tokens or more, such as a full transcript: documents above the query, instructions and examples. Copy this into a new chat inside the project:
| Drafting prompt |
|---|
| <notes>[transcript, voice memo text or bullet notes for this one post]</notes> |
| <examples> <example>[post 1]</example> <example>[post 2]</example> [3 to 8 more] </examples> |
| Write one LinkedIn post from my notes in the voice of the examples. Follow my voice rules. Use only what is in the notes. Give me two different first lines to choose from. |
The notes do the real work. McNulty uploaded a transcript of a podcast of about 90 minutes. Claude found about 60 possible posts in it and narrowed them to about 30 that didn't overlap, then drafted each one. McNulty still edited every draft, writing that Claude likes to use five words where two would do, and reported that the edited posts performed on par with fully manual ones. That is one person's self-reported result.
No podcast? Record yourself for ten minutes answering a question a customer asked you this week. Our guide to finding post ideas in sales calls, your inbox and your own decisions lists more sources. LinkedIn posts can be up to 3,000 characters, so the 1,300-character rule in the template leaves room for your edits.
How do you correct drafts so the next one is better?
Keep a correction log in the project and turn each repeated correction into a rule. A vague complaint teaches nothing. ContentIn's advice is to name the problem: the draft is too formal, or it opens in a way you never would.
Log the fix, not just the draft
One line per correction: date, what the draft did, what you do instead. For example: "24 Oct: opened with a question; I open with a number."
Save the log where the next chat can see it
In Claude, add the log to project knowledge. In ChatGPT, upload it as a project file. Either way, later chats in the project can draw on it.
Promote repeats to rules
When the same fix shows up three times, add it to the voice rules and delete it from the log.
Swap in your newest posts
Every month, replace your two oldest examples with two recent posts you edited heavily. They show how you write now.
Run a blind test
Mix three of your old posts with three edited drafts and ask a colleague which are yours. If they can't tell, the setup works.
What should you never hand over to the AI?
Your opinions, your experience and your numbers. The AI can shape them into a post, but it can't supply them. LinkedIn's own description of AI slop is content that "may sound polished on the surface but lacks a clear point of view, unique perspective, or substance". A voice setup fixes the sound. Only your notes fix the substance.
- Read every draft before it goes out. LinkedIn's help page says you are ultimately responsible for everything you post.
- Check every name and figure against your notes. An [ASK ME] left in a draft is a question you still owe the AI.
- Add one detail only you know. A customer's exact words, a number from your own dashboard, or a decision you got wrong. No model has seen it before.
If you'd rather pay someone to do all of this, our comparison of LinkedIn ghostwriters and AI tools sets out the costs and trade-offs. Siela's Social Media agent posts in your voice at the pace you set, and a second AI reviewer checks every post before it goes out. You can choose "show me every post first" or "post without asking" for each account.
Frequently asked questions
How many of my posts should I give ChatGPT to copy my writing style?
Five to ten is a practical start. Anthropic's prompting guide recommends 3 to 5 varied examples for best results, ContentIn suggests five to ten of your strongest LinkedIn posts, and one documented experiment used 25. Pick posts that sound like you rather than your most popular ones, and include at least two of each format you use.
Will LinkedIn reduce the reach of posts written with ChatGPT?
LinkedIn says its focus is whether content adds value, not how it was made. AI-assisted posts are welcome when they reflect a real person's perspective, experience or expertise. Content that appears AI-generated and lacks clear perspective is less likely to be widely distributed beyond your immediate network, according to its May 2026 announcement.
Should I use ChatGPT custom instructions or a project for LinkedIn posts?
Use a project. Custom instructions apply to every chat, and free accounts get only 1,500 characters for them. ChatGPT project instructions apply only inside the project and override your global custom instructions, so your LinkedIn voice rules stay separate from how you use ChatGPT for everything else, and your samples sit in the same place.
Do I have to say a LinkedIn post was written with AI?
LinkedIn's help page recommends letting others know if you relied heavily on AI to create or modify a post, unless that is obvious from the context. It also urges you to review, edit and approve any content AI helped with, and says you are responsible for everything you post. This is guidance from LinkedIn's help centre, not legal advice.
Why does Claude forget my corrections between chats?
Claude's help centre says context is not shared across chats within a project unless the information is added to the project knowledge base. Each project also has its own memory, but that is less explicit. Keep a short correction log as a file in project knowledge, update it after each edit, and turn any repeated correction into a rule.
Sources
- LinkedIn Pressroom: Keeping conversations real on LinkedIn (first published by Laura Lorenzetti on 20 May 2026) (published 4 Jun 2026)
- LinkedIn Help: Best practices for content created with the help of AI (checked 24 Oct 2026)
- LinkedIn Help: Share posts on LinkedIn (3,000-character limit) (checked 24 Oct 2026)
- OpenAI Help Center: ChatGPT Custom Instructions (checked 24 Oct 2026)
- OpenAI Help Center: Projects in ChatGPT (checked 24 Oct 2026)
- Claude Help Center: Understanding Claude's personalization features (checked 24 Oct 2026)
- Claude Help Center: How can I create and manage projects? (checked 24 Oct 2026)
- Anthropic: Prompting best practices (Claude docs) (checked 24 Oct 2026)
- Brendan McNulty, 52 AI experiments: Week 22, using AI to clone my LinkedIn voice (published 28 May 2025)
- ContentIn: How to train AI to write in your voice on LinkedIn (published 17 Nov 2024)