I have ADHD. In most parts of my life I’d call it a superpower. Writing is not one of them.
For a long time the pattern was the same. I’d have an idea worth writing about, and either I wouldn’t start or I’d start and never finish. The problem was never a lack of ideas or opinions. Ask me a question and I can usually answer it well. Where I fall apart is taking all those answers and putting them together into something organized that someone else would want to read.
So when I built kevinmazur.dev, I didn’t want a traditional blog or another system I’d constantly have to maintain. I wanted a way to get from an idea to a published piece without depending on the exact skill I’m worst at.
Two agents instead of one
The site was built with AI assistance, but that isn’t the interesting part anymore. The interesting part is the workflow around it, which uses two AI agents.
The first is Cooper, named after Anderson Cooper, because he’s a great interviewer and journalist. Cooper is my writing partner: interviewer, editor, and ghostwriter. I don’t ask it to write an article about a topic. I give it the topic and it interviews me. It asks one question at a time, follows the threads worth following, and pushes back when an answer is vague or a little too convenient. Only after that does it help turn the conversation into an article.
The second is Gutenberg, named for the printing press guy, for obvious reasons. Gutenberg takes a finished article and gets it published on the site. It deliberately doesn’t write or edit anything. If it finds a problem, it flags it and stops, because I would rather a post be delayed and correct than rushed out. So far it has published one post, the introduction to this site, and it did exactly what it was told.
That split wasn’t the original plan. I started out wanting one agent to do all of it. While I was working through the instructions, I realized that with a little tweaking the writing agent could be much broader than the website. It could help with email, reports, strategy documents, anything where my voice matters. Once Cooper grew into that, Gutenberg became a smaller, more precise tool. The separation came out of building it, not out of a principle I started with.
Whose thinking is it?
The obvious objection, and one Cooper raised while interviewing me for this piece, is that organizing ideas is part of thinking. If AI decides the order, what gets cut, and how one point leads to the next, am I really keeping the thinking human?
I think I am, but I don’t think the distinction is perfectly clean. Changing the order can change the emphasis, and leaving something out can change the argument. The ideas start with me, but I still have to decide whether the finished piece says what I mean. There have been times when Cooper didn’t get my meaning. When that happens, I explain it the way I would to a co-worker, and we go again.
Where I’ve landed, for now
I use AI to build software, and I don’t review every generated line myself. My process includes testing in development, running tests, and occasional review by others. That leaves a question I need to take seriously: how do I know that what it produces is sound? With writing, the responsibility is easier for me to see. I have to read the finished piece and decide whether it says what I mean. AI has been most useful here when it asks questions, organizes what I say, and handles the mechanics I’d otherwise use as an excuse not to finish.
If I had to put a rule on it, I’d say an AI agent has to start clearly defined, with the guardrails set, and then grow within them. Cooper is the example. It started as an article interviewer and is now a writing partner for almost everything I write. The guardrails didn’t go away. The role just got bigger inside them.
This article is the first real test of the whole thing. It came out of an interview, Cooper organized it, I’ll edit it, and Gutenberg will publish it. If you’re reading this, at least the second half worked.
