I wanted to write about an app I'd made, and my AI's first draft read like a summary of the product page. Accurate. Nobody's story.
Six questions in, it knew I owned the real instrument the app is based on. It knew I'd built the app once before and lost it. It knew I'd left out the advanced features on purpose. And it knew why I made it at all: so other people could feel the joy of the instrument without owning the hardware. None of that was findable. All of it went in the piece.
A list of ten questions is a form. One question at a time is a conversation, the way a good reporter does it. (I've been on both sides of that mic.)
So when I want something that sounds like me, I start like this:
Ask me one question at a time about this, and wait for my answer.
Follow what I actually said, not your list.
Stop when you have enough to write it, and tell me what you learned.
One question.
That wasn't its fault. Everything it knew was public, and the good parts weren't. So I asked it to interview me instead, one question at a time, and to let my answer pick the next question.