@VoltexGar: Boris Cherny, Head of Claude Code, Anthropic: "you can define a stop hook that's like, if the tests don't pass, keep go…

Boris Cherny, Head of Claude Code, Anthropic: "you can define a stop hook that's like, if the tests don't pass, keep going. essentially it's like you can just make the model keep going until the thing is done." that one config is the post's loop #1, straight from the person who built the tool. a stop hook fires every time the turn ends, so the work can't be called done until tests and lint go green. deterministic, zero effort, the loop most people skip. cherny describes the verification loop in his own words too: for big changes his team spawns subagents to find every issue, one checking claude.md compliance, one combing git history, one hunting bugs, then five more subagents whose only job is to kill the false positives. a fresh clean-context critic catches what the builder talked itself past. and the memory loop, in his words: if you see a mistake, you just tell claude to add it to claude.md, so next time it already knows. the correction outlives the session it happened in. four loops, none exotic. each is a few lines of config or one instruction. seniors out-ship everyone else because they run all four at once, not because their model is smarter.