Most advice about AI writing is a binary: use it and sound like everyone else, or don’t and keep your soul. That framing is wrong, and it’s costing people a lot of time.
The thing that determines whether AI output sounds like you isn’t the model. It’s what you put in. Hand it a topic and it fills the gap with the average of everything ever written on that topic, which is the definition of slop. Hand it your actual thinking and it has something real to work from.
Here’s how to do the second thing.
Start from your words, not a topic
The worst prompt is “write a post about consistency in content.” The model has nothing of yours, so it invents a plausible stranger and writes as them.
A better start is to dump what you actually think first. Three messy sentences about the time you posted daily for a month and what surprised you. Now the model has your example, your angle, and your surprise to build on. It’s editing you instead of replacing you.
The fastest way to produce that raw material is usually to say it, not type it. Talking is lower friction than writing, and it’s where your real phrasing lives. A few minutes of you explaining an idea out loud contains more usable voice than an hour of careful drafting, because you’re not self-censoring on the way out.
Make the model do the boring jobs, not the thinking
There’s a clean division of labor here. The model is good at the mechanical work and bad at the part that matters.
Good jobs to hand over:
- Cutting a rambling transcript down to the actual point
- Reshaping one idea into the format a given platform wants
- Fixing the flow when a paragraph is fighting you
- Catching where you repeated yourself
Jobs to keep:
- What the opinion actually is
- Which story to tell
- What you believe that the consensus doesn’t
- The final call on every sentence
When you keep the thinking and delegate the typing, the output stays yours. When you delegate the thinking, you get a stranger’s post with your name on it.
Edit for “would I say this”
Whatever a model hands back, run it through one filter before it goes out: read it aloud and cut anything you wouldn’t actually say.
This catches the slop reliably. The model will reach for “leverage” when you’d say “use.” It’ll add a tidy three-part summary you don’t need. It’ll smooth your sharp opinion into a safe one. Each of those is a small theft of voice, and reading aloud exposes them because your mouth refuses to say them.
This pass takes a couple of minutes and does more for authenticity than any prompt engineering.
Give it your back catalog if you can
If a tool lets you feed it things you’ve already written, do it. Your old posts are the best possible reference for your patterns: how you open, how long your sentences run, the jokes you make. A model with real examples of your writing imitates you, not a tone preset called “casual.”
The principle under all of this is simple. AI is a multiplier, and a multiplier does nothing to zero. If the input is a generic topic, you multiply nothing and get nothing. If the input is your real thinking in your real words, you multiply that, and you get more of yourself in less time. That’s the version worth using.
Talk through your ideas and publish posts that sound like you.
Claim a founding spot