Shipaton app "How We Work" misses ship after 2-week review stall

The project's timeline is the counterintuitive part. Coding itself stopped being the bottleneck once AI was involved, yet the whole build still took roughly two months, and the builder is explicit that it did not feel fast. What filled that time was judgement: deciding what to build, how far along things were, and how to keep the look consistent. Those are the things the builder moved out of the AI and into the repository as requirements, progress and rationale files.

Several of the individual techniques exist to serve that same goal. Rules were split from the history behind them after the AI started applying outdated rules it could not distinguish from current ones. A Kanban board held the AI's current task, the last instruction given and what remained, because across sessions the AI cannot carry over how far it got. An HTML requirements dashboard, generated after a single request, exposed contradictions between requirements that had been invisible in chat. Later, Mobbin examples replaced an instruction to make spacing and text sizes "nice", and the builder concluded that matching Figma exactly is not necessary when a real-world reference produces a better result.

The builder also found the AI absorbing the negative tone of failure notes kept in the repository, and had them moved to Obsidian. One unresolved question remains: how clean AI-written code needs to be. A deletion confirmation message pulls the word "delete" out of the message text rather than holding it as a variable, and while it works, only a second AI's code review caught it, and it was not fixed because the deadline was close.