A static portfolio has exactly one job — presenting information. I hit the same wall three times trying to make this one do more, and each time the fix looked the same before I'd even planned it that way.

The first cover-letter generator worked in an afternoon and was wrong in ways I didn't notice until much later. This is the story of the two separate times it broke, and what each one taught me about not trusting a model with the truth.

Giving an AI agent real access to a repository is fine, until you can no longer tell whether it acted or just proposed. Here's how I built the scaffolding that keeps that line visible — and where I got it wrong before I got it right.

Every time I added a new Claude call site, I wrote the same five lines of plumbing again and told myself it was fine because the feature was different. By the fourth time, it wasn't fine anymore — and the fix wasn't "write a better prompt."

`structuredGenerate()` fixed how a Claude call is shaped. It said nothing about what gets fed into one — and that turned out to be the harder problem, one I'd already half-solved by accident in the tables I'd built for an entirely different reason."

Structured data tells you what's true. It doesn't tell you how much of it to show a model, or in what order — and the two times I got that wrong looked nothing alike on the surface, but broke for the exact same reason.

The photography page sat untouched for seven months while everything else on this site got smarter. This is the story of why it took that long, and what finally broke the standstill.

The photo is the one input in this entire system I don't fully control. This is the story of the prompt, the contract, and the architecture decision that let Claude look at it safely.

Search shipped, and returned nothing. I fixed that, and it started returning the wrong thing with total confidence. Both failures taught me not to trust a similarity score more than it's earned.

Cyclist should have found the bicycle photo. It didn't, and fixing that taught me my first threshold guess was wrong twice — once at 0.75, and once again at a single flat 0.55.

A ranked list of photos answers "which ones match." It doesn't answer "why do these belong together," and that turned out to be the more interesting question.

A percentage badge and a list of cited photo IDs both look like proof. Neither one is, on its own — and the deterministic checks that had to sit behind each of them turned out to be very different problems.

A stray query token, a 0.75 threshold that still wasn't strict enough, and the discovery that some questions don't need a probability at all — only a rule.

Exact-hex color matching was never going to work, no matter how the threshold was tuned. Fixing the comparison model — twice, in two unrelated domains — came before any number did.

A permanent sweep against real ground truth, an unambiguous elbow, and one honestly-reported gap that no threshold, however tuned, was ever going to close.