Thoughts

Half of every AI integration project is not the AI. It is convincing the existing system to expose its data cleanly enough that the AI has something useful to work with. The model is the easy part.

Reading scripture in the morning and reading documentation in the afternoon. Both are practices of slow attention. Both reward returning to passages you thought you already understood.

The teams shipping useful AI features in 2026 are the ones who treat evaluation as a primary engineering discipline, not an afterthought. If you cannot tell whether your AI is getting better, you cannot make it better.

Spent the morning rewriting an article from 2022 about microservices. What was confident advice three years ago reads like overcorrection now. Worth doing this once a year. Past me is a useful teacher of present me's blind spots.

A tool that takes ten minutes to set up but saves five minutes a day pays for itself in two weeks. Most people stop at the setup cost. The math always favours the tool.