Three years

Three years ago this week, the most powerful tool in my studio was a text box that answered one question at a time, and I thought that was miraculous. It was. I just had no idea how small a unit of work that would turn out to be.

That's the measure I keep coming back to. Not what the machines can do — that headline changes every quarter — but the size of the job I can hand one and walk away from. December 2022: a sentence, a paragraph, a snippet I'd paste somewhere by hand. December 2025: a feature, scoped overnight, while a fleet of agents holds the thread for hours and I read the work instead of writing it.

The interesting part is what didn't change. When the machine got fast and cheap and abundant, the cleverness stopped being the hard problem. The hard problems were the patient ones. Will I trust this piece tomorrow, or will some agent quietly rebuild it with a new flaw? Whose context is this, and who decides when it's handed over?

That's where the studio's year went. Ctrl AI for the first question, HumanRepo for the second. I won't recap them here — they each have their own week coming. What matters tonight is that both started as the same itch: the loud layer this year was the models, and the loud layer is rarely the one that lasts.

I've shipped through enough cycles to distrust my own excitement and still feel it. Three years in, the excitement has just moved underneath the surface, to plumbing.

So next year is for building in the open. Fewer demos of what the machine can do. More careful work on what it's allowed to stand on, where anyone watching can check my seams.