A while ago I wrote about a feature-workflow plugin that added structure to how I work with Claude Code (the later update covers where it went): capture an idea, plan it, implement it, ship it. I've kept moving more of the process out of the conversation and into tools since then. muda does it for CI and deploy pipelines, and since mid-October I've planned most of my coding work in diazo, where the agent works through a page in my browser. It writes down what it understood, puts each open question in front of me as options with a recommendation, and then writes a spec and a plan that I approve before anything is built.
Both processes have felt like Claude does noticeably better work than without, and adding new features takes less of my own time, but I wanted to see what the data said. I had session logs from the weeks before diazo and the weeks after, so I went back through them to check. The logs back up the first impression: the agent made far fewer of the mistakes I have to stop and correct, and it used fewer tokens for the code it shipped. The second is weaker: my estimated time per merged PR fell only a little. I also catch more broken code than before, though most of it now turns up at a test step before anything merges.