I directed. I didn’t type.

Three side projects, six months, zero lines of code in my own hand. The surprise was not how much AI could build. It was what being able to build anything did to me.

Over the past six months I built three side projects, and I wrote none of the code in any of them. I did not type the specifications either. I assembled those with the model, from my own intent, the prototypes and art styles I chose, and the architecture I had in mind. I directed; I did not type.

That an AI can write the code, and that the work moves from writing to deciding, is by now the least interesting thing you can say about any of this. It is true, and everyone has said it. What is worth writing down is what happens after the building becomes easy, because that is where the surprises were, and they were not the ones I expected.

None of these began as “let me try AI.” Each was a problem I kept hitting and could not find a tool for: a way to ask questions of sensitive data without ever showing the data to a model; a way to read the world’s news without an algorithm deciding what should matter to me; and the one I care about most, an attempt to make a stranger on the other side of the world feel like a person instead of a category. Three problems. The AI was only ever the means.

Restraint became the scarce skill

AI makes it practical for one person to build what used to take a team. That sounds like pure upside until you live inside it. The risk is no longer that you cannot build enough. It is that you build far too much. When one more feature costs almost nothing, the discipline to not add it becomes the rarest thing you own.

The Front Page, the news project, is two Python scripts and a static page on purpose. Its entire value is that it is simple and shows you the world with no algorithm in the middle, and it would have been effortless to bury that under features I could ship in a weekend. Hermetic, the data tool, is the opposite: a heavy stack, because that problem needs one. The lesson lives between them. Match the machinery to the problem, not to what you are now capable of. AI removed the friction that used to stop me from over-building. The brakes had to come from me.

Generic is free; specific is the work

It is not enough to build less. What you do build has to be specific enough to be true, and generic is the direction a model drifts by default. Ask for something in general and you get something that sounds right and says nothing. The real thing lives in details a model will never volunteer, the ones you have to know to ask for.

That is the whole premise of the third project, Behind Every Them. It begins with a question: what if the person you think of as “them” is just you, somewhere else? You start at a globe and zoom until you are standing inside one family’s ordinary week, what they worry about, what they spend, what they cook on a Tuesday. Then it shows you a family on the far side of the world lying awake over the same things you do.

It only works on specifics. “They eat bread” is a stereotype wearing a fact’s clothing. The physics textbook a student saves up for, or the line item for generator fuel that quietly reorganizes a household’s month, those read as a person. Getting there meant pushing past the model’s first fluent generic answer, again and again, until the detail was specific enough to be someone’s real life. Specificity is the difference between a category and a person. AI hands you the generic version in an afternoon. The specific one is still yours to earn.

What I keep, and what I give away

The same instinct decided what I shared. I open-sourced one of the three and kept the other two close, and not for the reason people assume. The two I kept private are personal. I want them to stay scoped to what I have in mind, not pulled toward whatever the internet would turn them into. Hermetic is the opposite. I released it because I want people to take it somewhere I could never reach on my own. One kind of work you build to keep. The other you build to hand away. Knowing which is which is its own decision, and the model has no opinion on it.

A model can build any of these three. It cannot decide they are worth building, or which detail makes them real, or which to keep and which to give away. The question I open a project with is no longer “can I build this.” It is:

“should this exist, and am I the right person, for the right reason, to build it.”