How to make ChatGPT sound like you

ChatGPT writes a decent first draft. Getting it to sound like you is the hard part. Here's why it doesn't by default, how to fix it by hand, and where the manual way stops working.

Why AI drafts don't sound like you

People rarely spot AI writing from one word. They spot the pattern. Every sentence is about the same length. Every paragraph has the same tidy shape. There's a friendly opener, a polished sign-off, and nothing rough anywhere. Your own writing has habits that break that pattern. The AI's doesn't.

There's research on this. A study in Nature Human Behaviour looked at more than 880,000 texts and found that rewriting with an LLM cut the variation in writing complexity by 21 to 50%. Everyone's text drifts toward the same middle.

Fixing it by hand works less well than you'd think. In an ACL 2026 study, 81 people edited AI drafts to make them sound like themselves. Their edited versions still measured closer to the AI's style than to their own writing. The writers thought the edits had worked.

And readers notice. An ACL 2025 study found that people who use LLMs a lot are very good at spotting AI text. A majority vote of five of them got only 1 of 300 articles wrong. That beat most of the detectors the researchers tested.

The cleanup ritual

If you use ChatGPT to write, you probably already do this. Lots of people do. Pew found that 52% of workers who use AI chatbots at work use them to edit writing. In a Muck Rack survey reported by PRSA, 98% of PR pros who use AI said they always or often edit what it gives them.

The usual pass looks like this:

• Delete the opener. "I hope this finds you well" and "Great question!" go first.
• Cut the apology. "I sincerely apologize for any inconvenience" becomes "sorry", or nothing.
• Shorten the sentences. Split the long ones and drop the filler clauses.
• Tone down the enthusiasm. Most people aren't thrilled about a meeting moving to Thursday.
• Fix the sign-off. Use the one you actually use.

It works. But you do it every single time, and the result still reads like edited AI. That's what the ACL 2026 study found.

The manual method

You can get closer by telling ChatGPT how you write. Setup takes about 20 minutes.

1. Write down your patterns

Open a few things you wrote without AI: emails, Slack messages, a doc. Note:

• Sentence length. Short and blunt, or long with lots of commas?
• How you open. "Hey Sam," "Hi all," or straight into it?
• How you sign off. "Thanks," "Cheers," your initial, or nothing at all?
• Words you never use. Maybe "delve," "leverage" or "excited to share."
• Em dashes. Do you use them, or never?
• Other punctuation. Exclamation marks, semicolons, ellipses.
• Capitalization. All lowercase in chat? Capital letters everywhere?

2. Pick 3 to 5 real samples

Choose things you wrote yourself that are close to what you want ChatGPT to write. Work emails for work emails. Leave out anything AI already touched.

3. Give ChatGPT this prompt

I want you to write in my voice. Here's how I write:

- Sentence length: [short and direct / long with lots of commas]
- How I open: [e.g. "Hey Sam," or straight into the point]
- How I sign off: [e.g. "Thanks," or just my initial]
- Words I never use: [e.g. delve, leverage, excited to share]
- Em dashes: [I use them / I never use them]
- Other punctuation: [e.g. no exclamation marks, no semicolons]
- Capitalization: [e.g. all lowercase in chat]

Here are some things I wrote myself. Match their rhythm, tone and word choice. Don't copy their content.

Sample 1:
[paste]

Sample 2:
[paste]

Sample 3:
[paste]

Now write this in my voice. Keep every fact, name, number and date exactly as given. Don't add an opener, an apology or a summary unless my samples have one.

[what you want written, or the draft to rewrite]

Paste it at the start of a chat. You can also put your list of patterns in ChatGPT's custom instructions, so every chat starts with it. The samples are usually too long for that, so keep them in the chat or in a project.

Where it breaks

• It drifts. Longer pieces start in your voice and slide back to ChatGPT's defaults a few paragraphs in. Long chats do the same.
• You redo it in every tool. Claude, Gemini, Copilot and Cursor each need their own copy of the setup, and you keep them in sync by hand.
• Your list is a guess. Most people don't know their own habits that well. You'll write "I keep it short" and miss that you never start a sentence with "I".
• You still check every line, because it slips in phrases you'd never use.

The automatic version

This is the problem Thirdraft solves. You add a few things you've written. Thirdraft learns your patterns from them and rewrites any AI draft in your voice. Names, numbers and dates stay as written, and every change is marked so you can see what moved.

It works on the website and on your phone, in Chrome right where you type, and inside Claude, Claude Code and Cursor through the Thirdraft MCP server. Your voice lives in one account, so you set it up once. If you've been comparing tools, here's how Thirdraft differs from AI humanizers.

Sources

• Nature Human Behaviour: The shrinking landscape of linguistic diversity in the age of large language models (2026)
• ACL 2026: Can You Make It Sound Like You? Post-Editing LLM-Generated Text for Personal Style
• ACL 2025: People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text
• Pew Research Center: Workers' experience with AI chatbots in their jobs (2025)
• PRSA: Majority of PR pros using AI, survey finds (Muck Rack, 2026)

Questions