Claude Code can build a feature in minutes. The problem isn't speed — it's that "finished" usually means the code compiled and the tests it wrote itself went green. Nobody challenged the idea, nobody reviewed the diff with a critical eye, and nobody actually used the thing.
I wanted the opposite of a yes-man. So I built aSPARK: an AI product team for Claude Code that runs a gated delivery loop, where each step has to earn its way to the next one. This post is about the one decision that changed how much I trust the output — making QA click through the real app in a real browser — and what enforcing quality gates on an AI coding agent actually feels like in practice.
Top comments (0)