用 Jev 的 Choice 决策做收件箱分流，再交给 Claude
把来信先交给 Jev 做「进哪个队列、要不要回复、信息够不够」的选择题，常规件自动分流，复杂投诉才转给 Claude 处理。
STOP MAKING CLAUDE READ THE WHOLE DAMN INBOX. A newsletter. A duplicate lead. A customer saying “thanks.” A support ticket with three different problems in it. Sending all four straight to your most expensive model is a strange way to spend money. This is where Jev starts making sense. Give it the context and a small set of decisions: which queue, does this need a reply, is there enough information to move forward? For routing, a Choice answer gives you the selected option and probabilities across the options. Your code uses that result to choose the next step. An obvious newsletter goes into the newsletter queue. A messy customer complaint gets passed to Claude with the original message and the routing result. Claude gets something worth working on instead of being asked to explain, again, that a receipt is a receipt. incoming data → Jev → routine route / Claude / human review That same split fits support queues, lead qualification, document sorting and agent handoffs. You define the choices and what each one is allowed to trigger. Claude is available for the cases that need reasoning, investigation or an actual written answer. Check the cutoff against examples where you already know the right answer. A confident answer can still be wrong, so log the mistakes and give ambiguous cases somewhere to go. There’s no universal “90%” switch that makes every workflow reliable. Start with one queue you’re tired of checking. Give the difficult cases your best model’s attention. The newsletter has had enough. The article below goes deeper into wiring that decision layer into an agent.
这条是引用着另一条写的：x.com/beaver_0x/status/2105662726655590821