Using writing history to make voice drafts sound like me

I use voice a lot with AI. Most of the time I’m not sitting down and carefully writing a prompt, I’m just talking through whatever is in my head, brain dumping the context, what I’m trying to say, and what I need it to do. Then I use AI to clean that up and turn it into an email, post, message, plan, or whatever I’m working on.

One of the problems with that workflow has always been getting the finished writing to actually sound like me.

So I had Codex go through all of my sent emails in Gmail and build a writing profile based on how I actually communicate. It went through all sent messages, filtered out the automated unsubscribe stuff, signatures, forwarded messages, code dumps, etc., and looked at how I structure sentences, how formal I am, how I make requests, the words I tend to use, and how my longer messages differ from quick replies.

Then we turned all of that into instructions it can use when it takes one of my voice brain dumps and turns it into something I’m actually going to send or post.

That’s a lot more useful than repeatedly telling it to “make this sound like me,” because now there’s actual writing history behind that instruction. It even picked up on some of the things that drive me crazy in AI writing, like a bunch of short dramatic paragraphs, overly polished corporate language, unnecessary em dashes, and all the “This isn’t X. It’s Y.” stuff.

You could do the same thing with a Facebook / LinkedIn / X export of your old posts too. Email gives it a really good picture of how you communicate directly with people, and your social history can give it more examples of how you naturally write when you’re talking publicly and casually.

For me, this fits really well with using voice as the input. I can focus on getting the thoughts out of my head without worrying about how I’m wording everything, and then AI has a much better idea of how to turn those thoughts into something that still sounds like I wrote it.