Three lines, said once, and every draft comes out sounding like you instead of a press release. Setup time: about 20 minutes, and most of that is finding old emails.
Wikipedia's editors got so tired of AI-generated junk that they published a field guide for spotting it. This playbook flips that guide into a ban list, pairs it with samples of your real writing, and locks both into ChatGPT's memory so the setup survives every future chat. You never repeat it again.
Before you touch ChatGPT, collect 5 to 10 samples of writing that is genuinely yours. Sent emails are the best source because nobody performs in an email. Slack messages, old blog posts, and text threads work too. Ideally grab stuff you wrote before you started using AI at all.
Two rules for picking samples. First, they should sound like you on a normal day, not your most polished work. Second, strip out anything sensitive: client names, numbers, addresses. You are teaching rhythm, not sharing secrets. Paste them all into one doc so they're ready.
Open a fresh ChatGPT conversation and paste this:
The page is real. Search "Wikipedia Signs of AI writing" and read it yourself; it is the sharpest catalog of AI tells anywhere because it was written by people who delete AI slop for fun. Asking ChatGPT to list the patterns back is your verification that it actually processed the list instead of nodding along.
Same conversation. Paste this, followed by your samples from Step 1:
The summary it produces is the important part. Read it. If it says you write short punchy sentences and you actually ramble, correct it right there: "No, I use long sentences with commas, fix rule 3." You are editing your own style profile, and this is the only time you'll ever have to do it.
This is the line everybody skips, and it's the one that makes the whole thing stick. Without it, your training dies the second you close the chat.
ChatGPT will confirm it updated its memory. To double-check, go to Settings, then Personalization, then Memory, and you'll see the saved entries. Memory needs to be switched on for this to work; it usually is by default, but if that toggle is off, flip it.
Open a brand new chat. This matters: a new chat proves the memory took, because nothing from your setup conversation carries over except what got saved.
Read the output next to one of your real emails. Same greeting style? Same sentence length? No banned words creeping in? If something is off, say exactly what and finish with "update your memory with that." Every correction compounds. Two or three rounds of this and the drafts stop needing edits at all.
The same three lines work in the other major tools, they just save differently.
Claude: create a Project and paste your ban list plus style summary into the Project's custom instructions. Everything in that Project follows your voice. Claude's Styles feature is even more direct: choose "Create custom style," upload your writing samples, and Claude builds a reusable style you can switch on in any chat.
Gemini: go to Settings, then "Saved Info," and paste your ban list and style rules there. Gemini applies saved info across future conversations, same idea as ChatGPT's memory.
Keep your master doc from Step 1 with the samples, the ban list, and the final style summary. New tool comes out next year? Ten-minute setup instead of starting from zero.
The whole trick in one sentence: the fix was never teaching AI to write better. It was telling it exactly what to knock off, then showing it what you sound like instead.
This playbook is one piece of the AI Automation Starter Stack: 50+ free playbooks for putting AI to work on the stuff you actually do every day.
Get all 50+ free playbooks