Jev 打分钩子的上手教程
十分钟接入 typesafe-sdk,用 Choice/Score/Noul 一次评 40 个钩子,只放 3 个进真模型,附升级顺序。
How to use Jev, and where it gives you the most advantage: setup takes 10 minutes: join the waitlist, people are getting access same day 🔗https://typesafe.ai 1. install the sdk pip install typesafe-sdk 2. write the questions once, reuse them forever from typesafe_sdk import Choice, Noul, Score, TypeSafeClient client = TypeSafeClient() r = client.evaluate( state=draft_hook, questions={ "angle": Choice( instructions="what kind of hook is this", criteria={ "number": "leads with a specific figure", "contrarian": "argues against the common take", "howto": "promises a procedure", "news": "reports something that happened", }, ), "specificity": Score( instructions="how concrete is the claim", legend={0: "vague", 1: "one real detail", 2: "checkable numbers"}, ), "seen_before": Noul(instructions="does this read like every other post"), }, ) 3. read three numbers instead of a paragraph no prose to parse. angle comes back as a label with probabilities, specificity as a number on your own scale, seen_before as a raw 0 to 1. now the part nobody is posting: questions in one call run in parallel and barely move the latency, and output tokens are free. so ask every question you might need, including the ones you'll throw away. i score 40 hooks at once and only three reach a real model. the writing budget goes to the survivors. here is the upgrade order: 1. write the criteria like a rubric, not a prompt each option gets a description. that description is what the model matches against, so vague options produce vague labels. 2. keep option lists closed Choice takes 2 to 255 options and validates locally, so a bad list costs you nothing. build the list in code, from your own taxonomy, then let it pick. 3. threshold on confidence under 0.5 escalate to a big model or a human. 0.85+ before anything irreversible, which for content means anything that ships. 4. score the whole backlog, not the next post the cheap move is running it on one draft. the real one is running it across everything you've ever published and finding what your best posts share. 5. put it before the expensive step the frontier model only sees work that already passed. I'll leave the full doc in comments, check it out and read the article on full Jev setup from scratch
这条是引用着另一条写的:x.com/zodchiii/status/2101243146596384854