Newsletter
One decision from a recent build, including where our assumptions were wrong, and where we have room to plan ahead.
AI-generated code rarely fails outright, it just looks reasonable enough to merge. That's a harder problem than bad code ever was.
Every dashboard said the app was healthy. Users still called it slow. The fix wasn't performance, it was perceived responsiveness.
Bigger context windows made retrieval lazy. Curated, smaller context beat sending everything, every time we tested it.
Reserved instances and rightsizing barely moved the needle. The real cost was hiding in services that quietly polled every few minutes doing nothing.
Token usage went down, cost went down, satisfaction stayed flat. The real signal was how often the assistant admitted it didn't know.
The worst outages rarely came from one complicated feature. They came from complicated releases, five things changing at once.
A design system with a component for every edge case left nobody sure which one to use. We cut 40 of them and confidence went up.