What the AI didn't do

I built three products in about a year while working full time. That is not a thing I could have done in 2023, and pretending otherwise would be silly. All three were built with Claude Code.

So here is the specific version, because the vague version helps nobody.

What it did. Component scaffolding. TypeScript plumbing. The fourth identical card variant. Migrations. Reading a stack trace at one in the morning faster than I can. The parts of building that are typing rather than thinking — which, on a solo project, is most of the hours.

What it didn’t do. Decide that discovery on Então? Comer! should return three results instead of forty, because a list is a way of declining to answer the question. Decide to run the quiz combinations against the database rather than looking at them, which is the only reason I found that 44 of 288 paths returned nothing. Notice that the most important control in the product was 174 pixels below the fold. Write the verification script before rewriting 160 catalogue fields, which is the only reason I found the three details I had invented. Find the stored XSS in DesignGrounds’ progress import, which came from deciding to audit my own code as if a stranger had written it.

Every one of those is a judgement about what the product is for, made by someone holding the whole problem in their head. The model was in the room for all of them and led none of them.

The honest summary is that AI moved my ceiling, not my judgement. It let one designer ship three products. It did not make any of them good, and on the two occasions I let it drive a structural decision, I rebuilt that part later.

I mention this at the top of my portfolio rather than waiting to be asked. If the tooling is going to be visible in the work, I would rather be the one who points at it.