Jev 从 0 到进阶的 20 条使用技巧

Movez @0xMovez原推发布于 2026.09.19

在 X 上打开原推

20 TIPS FOR USING JEV FROM 0 TO PRO TypeSafe just shipped the first model built for machines, not for people. Here are my top 20 tips on how to actually use Jev: 1. stop treating it like an LLM. Jev doesn't write, it decides. state in, typed decision out. 2. three primitives: Choice picks from a list, Score places it on a rubric, Noul returns a 0-1 probability. 3. the architecture: LLM generates, Jev decides, code controls, humans catch the uncertain cases. 4. skip personas and preambles. state + one atomic question + exact criteria. 5. descriptive criteria beat labels. "technical = broken behavior or integration failure" beats "pick the right team". 6. one question, one judgment. "valuable, urgent and likely to buy?" is three questions you combine in code. 7. never ask it to explain itself. decision, probability, confidence. that's the contract. 8. keep state clean: the object, the context to read it, the facts that change the call. nothing else. 9. 64k limit, but accuracy shifts as state grows. the smaller useful state is the upgrade. 10. batch your questions. one call can judge 13 dimensions in parallel and barely move latency. 11. ask everything you might need, even what you'll throw away. output tokens are free. 12. gate on confidence, not the answer. act above 0.85, escalate 0.55-0.85, human below 0.55. 13. tune those thresholds on your own data. numbers from a post are a starting point, not a config. 14. every uncertain answer needs a route. low confidence treated as normal is how bad decisions ship. 15. never let it invent options. build the candidate list in code, then let it pick. 16. keep arithmetic and hard rules in code. Jev judges meaning, your program does the math. 17. pin the version. jev-latest to track stable, jev-1.13.0 once your thresholds are tuned. 18. log model ID, probabilities, confidence, route, outcome. review false positives, retune. 19. put it in the loop, not next to it: route before the call, gate the tool, judge the output. 20. the five money workflows: universal verifier, support triage, lead scoring, model routing, giant dataset jobs. setup: waitlist at typesafe .ai, then read how to build the fastest ai agents systems with Jev in article below.

这条是引用着另一条写的：x.com/0xmovez/status/2101007482919227841