firehose/app/priv/blog/engineering/2026/07-13-decide-the-size-of-your-step-2026-edition.md

%{ title: "Decide the size of your step, 2026 edition", author: "Willem van den Ende", tags: ~w(AI coding-agents practice), description: "When introducing TDD I always say you have to be able to move forward in really small steps, even when you take bigger steps. That way you always have something to fall back on.", published: false }

This post has been typed by hand, and describes my current way of dealing with taking much larger steps using coding agents. I enjoy building quickly, seeing usable software take shape. I can do that fairly safely using synthetic TDD, but have found that the resulting system is not always as coherent as I would like it to be. Look carefully in the screenshots in the scheduled posts release notes and you will notice that there is a Dashboard showing scheduled posts, but there is no navigation to Dashboard. Claude Code had not needed it to produce the demo. A coding agent can remember many more URLs than I can, and it can find them if the navigation is not there. A similar thing happened in the Federated Business Intelligence App. No navigation links, but a slick demo.

WeReview 2 was one of the first things I did mostly hands-off, over a year ago. I had pointed Claude Code at the source for WeReview 1, asked it to analyse what was there. I did mention my overall goals, broke down the work in parts, and was pleasantly surprised that I got most of WeReview 1 plus a fix for some UX annoyances in about five days of calendar time. It was half-term and I was also doing things with my family. This was one of the cases where I experienced AI whiplash.

When introducing TDD I always say you have to be able to move forward in really small steps, even when you take bigger steps. That way you always have something to fall back on., I have been mentioning this recently in QWAN's AI with TDD workshop, as well as in a keynote I developed for Philips.