Building an AI Writing System That Actually Sounds Like You

The first few times I asked AI to draft something for me, the result was competent and wrong. It was grammatically fine. It just didn’t sound like anyone I knew, least of all me.

It took me a while to see the problem clearly. I had never told the assistant how I write. I had only told it what I wanted written. So it fell back on the average of everything it had read, and the average sounds like a press release.

Describe the voice, not just the task

The fix started with a style guide for myself. Not a list of adjectives like “warm” and “professional,” which every style guide says and no assistant can act on. I needed specifics. How long are my sentences, and how much do they vary? Which words do I never use? Do I use contractions? How do I open an email to someone I’ve known for twenty years compared with someone I’ve never met?

That last question led to voice profiles. I don’t write to a customer the way I write to one of my kids, and the assistant shouldn’t either. A short profile for each relationship tells it how to adjust without abandoning the underlying voice.

Name the habits you want gone

Some of the most useful guidance is negative. AI drafts have recognizable habits. They group things in threes when there are really two or four. They set up a contrast that wasn’t needed. They end paragraphs with a tidy line that sounds more certain than the evidence. They reach for big abstract nouns where a plain verb would do.

I wrote those habits down as things to check for. I’m careful about one point, though. None of these patterns proves a sentence was written by AI, and plenty of good human writers use them on purpose. They’re editing clues. The question is always whether the sentence serves the reader.

Give it a workflow and a final check

A voice description isn’t enough on its own. I also wrote down how I want drafting to work. Settle who the reader is and what they need to do next, write a draft, then revise against the standards before anything is called finished. The last step is a short review checklist that covers truth first and style last.

Keep versions

My standards have changed several times as I’ve learned what the assistant gets wrong. Each version stays in the library beside the one that replaced it, with a short note about what changed. That history has been more useful than I expected, because it shows which rules actually fixed something.

The result isn’t a perfect imitation. I still edit. But I’m editing a draft that sounds like me, which is a very different job from rewriting one that doesn’t.