用 Jev 当路由器,把 Opus 5.5、GPT-6 Astra、Grok 编成一队
作者让 Jev 按任务难度分配模型:Opus 管架构推理、Astra 管草稿代码、Grok 管实时信息,称一周账单从 1940 美元降到 287 美元。
i finally figured out how to run dario amodei's opus 5.5, sam altman's gpt-6 astra and elon musk's grok as one team... and the boss is a model that costs four cents the trick: you never pick the model yourself. jev reads every task and hands it to whichever of the three actually deserves it, at the effort it actually needs. [here's how i'd wire it:] claude breaks the project into tasks. jev scores each one and assigns the model and the effort. each model only ever gets the work it wins. → opus 5.5, high: architecture, hard reasoning, the calls that need real judgment → gpt-6 astra, medium: drafts, code, the high-volume grind it cranks out fast → grok, medium: anything live, anything that needs what happened on x an hour ago → jev, router: reads scope, uncertainty and what failure costs, picks the model, never writes a word itself my first week on it: 4,812 tasks routed. jev benched opus on 81% of them, because opus kept writing essays for tasks that needed one line. the bill went from $1,940 to $287, and output went up. jev gets the task's scope, what is uncertain and the consequences of failure. it only picks from the models allowed for that task. your approval checkpoints stay in place. paste this into your next planning session: "use jev as the router in front of opus 5.5, gpt-6 astra and grok. break my request into tasks with clear owners, dependencies and a definition of done. for each task, have jev score reasoning depth, speed need and failure cost, then assign one model: opus 5.5 for deep reasoning, gpt-6 astra for fast drafting and code, grok for anything that needs live information. never route a task to a model it does not fit just to spread the load. show me the assignments before anything runs. after i approve, dispatch each task at its chosen effort. keep me on the line for anything irreversible. at the end, report which model did what, the total cost, and any task jev could not confidently route." steal this
这条是引用着另一条写的:「一篇上手 Jev 的完整构建指南:代码、架构与图」