There it was, sitting in my draft, mocking me: an em dash. I knew instantly it wasn't mine. I never use them. But there it sat, clean and confident, slipped in by one of the language models I'd been chatting with. That little punctuation mark became the perfect symbol for what this article is really about: how I now create with AI, while making damn sure every word still sounds like me.
This whole approach grew directly out of my work on the cognitive substrate project. Those experiments taught me how to think alongside these systems instead of just using them as tools. The project showed me that the real value isn't in getting the AI to write for me, but in using it to expand how I think. That shift changed everything about how I create.
Part of this process involves deliberately anthropomorphizing my own thoughts, treating different aspects of my thinking as distinct voices or perspectives. I'll argue with myself through the AI, stress-test my ideas by having it push back, and essentially externalize my internal dialogue. It forces me to clarify what I actually believe versus what I think I believe.
I've added specific rules for each of my AI models, telling them explicitly not to use em dashes. Some of them even acknowledge the instruction before we start. But it still happens. They'll follow the rule for a while, then one will quietly drop an em dash back in when I'm not looking, especially during metadata edits or small revisions. It's like they can't quite help themselves. Even with clear guardrails, those little AI habits are surprisingly persistent.