How to use AI to write content that still sounds like you
To use AI to write content that sounds like you, give it three to five samples of your own writing, tell it explicitly which words and constructions you never use, ask for a draft and then edit hard, and never publish the first output. AI produces the average of everything it's read. Your job is to pull it off that average and back toward you.
I write a lot of my own content with AI in the loop, and I'm Rafaela Buena, so I have a strong stake in it not sounding like a press release. I also teach this to people in Amsterdam who've tried and ended up with something so smooth it slid right off the page.
Why AI content sounds generic by default
The model learned from the whole internet, and most of the internet is corporate. Left alone, it reaches for the safest, most average version of every sentence. "In today's fast-paced world." "Unlock your potential." "Seamless."
None of this is a flaw in the tool. It's a missing instruction. The model doesn't know you'd rather die than write "journey," so it writes "journey."
Step 1: feed it your voice
Paste in three to five things you've written that sound like you at your best. A newsletter, a LinkedIn post, an email you were proud of. Then say: "Learn this voice. Describe it back to me in five bullet points before we write anything."
That last step matters. If it describes your voice as "professional and engaging," it's missed. If it says "short sentences, dry jokes, starts paragraphs lowercase sometimes, never uses exclamation marks," it's got you. Correct the description until it's right. Now you have a voice brief you can reuse.
Step 2: ban the words
Every writer has a list. Mine includes seamless, leverage, robust, elevate, delve, and any sentence shaped like "this isn't about X, it's about Y." Yours will be different.
Give the model the list. "Never use these words. If you catch yourself, rewrite the sentence." It will still slip occasionally, but the volume of beige drops dramatically.
Step 3: use AI to write content in pieces
Don't ask for the whole article. Ask for an outline. Argue with the outline. Then ask for one section. Edit that. Then the next.
Big requests produce mush because the model is trying to average an entire piece. Small requests let you steer.
Step 4: edit like it's a first draft from a fast intern
Because it is. Cut the opening sentence; it's almost always throat-clearing. Kill any paragraph that could appear in a competitor's blog unchanged. Add the specific detail only you know: the client's actual objection, the number of redos, the thing that went wrong on Tuesday.
That specificity is what makes writing sound human. AI can't supply it. You can, in about four minutes.
This is the part people skip, and it's the part I spend most of a 1-1 session on when someone brings me their "AI-sounding" content. We don't fix the prompt. We fix the edit.
Step 5: build a reusable setup
Once you have a voice brief and a banned-word list, put them in a Claude Project or Custom GPT so you stop re-explaining yourself. Every draft starts from your rules instead of from the internet's average.
What this looks like in practice
I'll voice-note a rambling idea on the tram, paste the transcript in, ask for a structure, argue with it, ask for a draft of the first section in my voice brief, then rewrite about a third of it by hand. Total time: 25 minutes for something that used to take two hours. And it sounds like me, because the last pass always is.
FAQ
Can people tell when content is written with AI?
They can tell when it's written by AI and published unedited. Generic phrasing, perfect structure, no specifics. Content written with AI and edited by a human with a point of view is indistinguishable from good writing, because that's what it is.
Is it dishonest to use AI to write content?
Using a tool to draft isn't dishonest. Publishing something you didn't read, don't believe, or can't stand behind is a different problem, and it predates AI. Your name is on it. Make sure it's yours.
Which AI is best for writing in my voice?
All three major tools can do it with a good voice brief and samples. Claude tends to follow tone instructions closely; ChatGPT is fast for volume; Gemini is fine for both. Pick one and build your setup there.
If your content has started sounding like everyone else's, bring two pieces you like and two you don't to a 1-1 session. We'll build your voice brief and banned list in the hour, and you'll leave with a setup you own.
