AI can work in parallel. I cannot.

Draft note. I am testing this idea in my work now, so my conclusion may change.

There are at least two ways to build something.

In the first, you do a rough pass across everything. You build the backend, frontend, onboarding, copy, emails, analytics, and all the other parts to a level where they mostly work. Once the whole thing exists, you return and improve each part.

In the second, you take one part and do it well. You stay with it until it is clear, useful, and complete enough. Then you move to the next part.

I have mostly been doing the first one while coding with AI. It feels natural because AI makes parallel work look cheap. I can ask it to build one feature, fix another page, investigate a bug, and rewrite some copy at the same time. The AI usually does fine on each task.

But I was still building mediocre results across all of them.

The problem was not always the AI's output. The problem was the attention required from me. Every part needed context. I had to decide what good meant, notice what was missing, give feedback, check how it connected to the rest of the product, and remember which compromises were meant to be temporary.

AI could work on many things in parallel. I could not judge many things well in parallel.

At the beginning, I could keep up. The project was exciting, my mind was fresh, and opening several paths felt like momentum. But as time passed, the unfinished decisions accumulated. Every task left a little state in my head. Switching between them became heavier. I became stressed more quickly, gave less precise feedback, and accepted work I would have rejected earlier.

The quality did not collapse all at once. It degraded quietly. One vague interaction stayed vague. One shortcut became part of the architecture. One piece of copy was good enough for now. Soon the whole product worked, but nothing in it felt carefully resolved.

I had treated AI's ability to generate in parallel as if it had increased my ability to pay attention in parallel. It had not.

Now I want to test the other way.

I want to choose one small, complete part of the product and stay with it until I understand it and feel it is good. Not perfect, and not endlessly polished, but good enough that I am not carrying a silent promise to fix it later. Then I can close that loop and move to the next one.

This may look slower because fewer things will be moving at once. But perhaps visible activity was the wrong thing to measure. If doing many things roughly creates a large second pass, then the speed was partly borrowed from the future. I still have to repay it, now with more context to recover and more mediocre decisions built on top of one another.

The bottleneck in AI-assisted coding may not be how much code the AI can produce. It may be how much care the human can sustain.

For now, I want to build at the speed of my attention: one thing done well, then the next.