TypeSafe 官方 Jev 提问设计手册，7 个模式
官方 playbook 把「写提示词」换成「设计问题」：一题一判断、字段路径、对比标准、留逃生口、批量扇出、置信度路由、分维度打分，附 5 条硬规则。
holy sh*t, TypeSafe AI published the official playbook for talking to Jev, and the first rule is: stop writing prompts Jev doesn't read prompts. it answers typed questions. their guide calls it question design, and I squeezed it into 7 patterns: 1. one judgment per question "is this spam?" hides six decisions behind one number. ask six narrow Nouls instead and weigh them in code 2. point at the field write http://support.tickets[0].message in the question. backticked paths kill the ambiguity 3. contrastive criteria give every option a "what" and a "not_for", so two similar answers stop blurring together 4. leave an escape hatch add "needs_review" to every Choice. without it Jev has to pick the nearest wrong option 5. speculative fan-out ask every question you might need in one request. their own cookbook: 13 questions in one call vs 13 calls, same answers, 12.2x cheaper and 10x faster 6. confidence-gated routing the answer says what. the confidence says whether you may act. below 0.8 it goes to a human 7. composite scoring never ask "rate this". score each dimension, keep the weights in your repo and 5 rules it will never break for you: > it reads your question literally > question IDs are for your code, the model never sees them > Noul at 0.5 means unsure, not medium > counting and date math belong in code > never ask it for the reply, it can't write one prompt engineering had a good run. this is question engineering full cheatsheet below, save it before your next build
这条是引用着另一条写的：x.com/n01ennn/status/2103542021071978601