Jev 入门解释：给选项打分而不是让模型思考

Termsheetinator @termsheetinator原推发布于 2026.09.20

在 X 上打开原推

Everyone is hyping up Jev and somehow making it sound 10x more complicated than it is. Here’s Jev for dummies: Jev basically does one thing extremely well, extremely fast for extremely cheap.. Your harness gives it information + choices. It tells your harness which choice is most likely correct. That’s it. - Normal LLM routing a support ticket: Ticket: “Hey, I was charged twice this month. Can someone fix this?” The LLM has to work through something like: “The customer mentions being charged twice. That sounds payment-related. Could this be general support? Maybe. Could it be account management? Probably not. Billing handles invoices, charges and refunds. So billing is most likely the correct department.” Then: Answer: Billing That whole process is overkill when the only choices are: > Sales > Billing > Technical Support > Account Management > Cancellations Jev is basically: Ticket: “I was charged twice this month.” And immediately scores the options: > Billing: 97% > Technical Support: 1% > Account Management: 1% > Cancellations: 0.7% > Sales: 0.3% > Route to Billing. LLMs can spend time “thinking” their way to a choice. Jev is built to just score the choices and pick one. When you have an AI agent doing that thousands or millions of times, that’s where it starts making a ton of sense. Think: “Is this company an ICP?” Yes / No “What type of reply is this?” Interested / Objection / OOO / Unsubscribe “Which campaign should this lead go into?” Campaign A / B / C “Which tool should the agent use next?” Tool A / B / C “Is this task finished?” Yes / No “Does this need a smarter model?” Yes / No It’s really good for: > classification > routing > scoring > ranking > yes/no decisions > choosing between known actions > validating information > deciding what an agent should do next Basically: Codex = does the work. Jev = helps Codex make a shit ton of tiny decisions while doing the work. Now install it. Sign up for the waiting list, I got my invite link within ~20hrs Install the official TypeSafe skill: npx skills add typesafe-ai/skills --skill typesafe-ai Select Codex when it asks which agent. Want it globally.. npx skills add typesafe-ai/skills --skill typesafe-ai -g Add your API key: TYPESAFE_API_KEY=your_key Tell Codex once: “Use the TypeSafe skill to find places in this project where Jev should replace brittle rules, parsing, classification, routing, scoring or closed-set decisions.” If you want Codex thinking about Jev automatically, add this to your AGENTS.md: “Use TypeSafe/Jev by default whenever a task involves classification, routing, ranking, scoring, validation, or choosing between known actions. Prefer Jev over a generative LLM for closed-set decisions.” Now you don’t need to type: “USE JEV” every 5 seconds. The skill teaches Codex how TypeSafe works, its question types, patterns and best practices. And your AGENTS.md tells Codex when you want it applied. That’s it, simple.