Dispatches

Not a feature comparison. A three-month field report on what actually changed how I write code — and what stayed noise. Includes the specific moment I stopped switching back.

AI code review is getting good at finding the wrong things. We ran it on three real PRs and compared the flags to what senior reviewers actually cared about.

The ecosystem has matured enough that the onboarding tax is real but finite. This is where we land after running it as a daily driver for eighteen months with a real client workload.

We ran Ollama with four models for six weeks on two machines. The gap between "it works" and "it's useful" is still significant — but it's closing in specific, predictable ways.

Nobody told me my shell history was worth organizing. It took losing a six-step deploy sequence to convince me. Here's the setup that finally made it searchable.

The resource usage was the last straw but not the first problem. A walkthrough of the replacement setup — colima, nerdctl, and what's still missing.

Not the productivity-science version. The actual cost in a two-person engineering team with four active services. What we measured, what we changed, and what didn't help.

A repeatable process for not accumulating tools you're not using. The actual questions I ask, the spreadsheet I keep, and what triggered the last three changes.