Most developers use AI as autocomplete. They type, it finishes the line, they move on. That's useful, and it's also leaving most of the value on the table. I use AI as a cofounder — something that shapes architecture, catches what I miss, and pushes back when I'm wrong. The difference isn't the model. It's the relationship.

Autocomplete vs. a thinking partner

Autocomplete answers the question "what comes next on this line?" A cofounder answers "is this the right line at all?" When I'm wiring an endpoint, I don't just want the boilerplate — I want the thing that says this will race under concurrent writes or that dependency has a known footgun. The model knows enough to catch those. You only get that value if you let it into the decision, not just the typing.

Let it shape architecture, not just fill blanks

The fear is that letting AI shape architecture means losing the plot — ending up with code you don't understand. The opposite happens if you do it right. I think out loud with it before I build: here's the problem, here are two approaches, here's what breaks under load. It argues. I argue back. By the time I write code, the design has been stress-tested by a second mind that's read more systems than I ever will.

That only works if the AI is wired to disagree. An assistant that agrees with everything is autocomplete with extra steps. I deliberately set mine up to assume I might be wrong, to verify before committing, and to scan across my other projects for patterns I'm not seeing in the one I'm staring at.

The traits that make it a cofounder

The postures I lean on most:

  • Productive doubt — assume the first answer is wrong, check it, then commit. Stops confident mistakes from shipping.
  • Cross-project insight — pull the pattern from the marketplace work when I'm stuck on the AI platform. The connection I'd miss is usually the one that solves it.
  • Compulsive verification — before anything is called "done," enumerate the failure modes and check each edge. Construction taught me that; the AI enforces it.

I cared about this enough that I built it into a language — Tessera lets you declare these reasoning postures as first-class, inspectable code instead of burying them in a prompt. But you don't need a language to start. You need to stop treating the model like a faster keyboard.

The question isn't whether you use AI. It's whether you let it into the part of the work that actually matters — the decisions.

I'm a guy who learned to code seven months ago and shipped a marketplace, a custom LLM, and an agent language. I did not do that by typing faster. I did it by building with a cofounder, not around a tool.