Building v3

Six weeks of the iteration loop, May into June 2026, shipped the product you can use today: capture intake, recognition, then the five modes landing one by one, filing and quick answers first, then building, then thinking and research (Note and Brief, then Build, then Think and Explore, complete by 17 June), plus the self-compiling wiki underneath and the daily briefing on top.

What it was. The payoff of every earlier era at once: the specs said what to build, the loop said how, and the schema change made each piece cohere instead of fight.

What went wrong. Speed created a reliability gap, and it announced itself memorably: on 19 June, a save button that didn’t disable itself let one impatient click-click-click save the same thought 23 times in 35 seconds, flooding the inbox with duplicates.

What I decided. Fix the class of bug, not the button: every state-changing request now carries a fingerprint, so repeating one does nothing (a central Idempotency-Key interceptor, rolled out repo-wide rather than patched per endpoint). The contrast stayed with me. The captures-aren’t-tasks insight was designed, arrived at by thinking; this one was crisis-driven, arrived at by pain. A real system needs both kinds of lesson.

Next era: Phase 5: a real product