Jev 创始人发布 12 页编码 agent 使用蓝图

Hanako @hanakoxbt原推发布于 2026.09.27

在 X 上打开原推

Jev Founder, Diogo Almeida, just released a 12-page PDF on how to use Jev with coding agents It is more useful than most paid AI courses: this is a 10-step blueprint on how to make Claude Code, Codex or any coding agent cheaper, faster and safer to leave running: step 1 → count the forks: most agent turns decide something, only a few actually write code step 2 → build the turn snapshot from git and the test runner, never from the agent's summary of its own work step 3 → route before the call: a Choice picks cheap, mid or frontier, and falls upward when unsure step 4 → guard every command: deny rules first in code, Jev only for what a regex cannot catch, "ask" instead of "block" step 5 → score the context: every tool output gets a 0 to 3 relevance score, survivors stay verbatim step 6 → check the exit: one Noul over fresh test output decides whether the agent is actually done step 7 → gate the merge: verdict, blast radius and needs-human in one call, hard rule checked first step 8 → cascade the tiers: the cheapest model tries first, only failures climb step 9 → run everything in shadow mode for a week and read your thresholds off your own receipts step 10 → keep the receipt: turn, question, probabilities, tier, decision, outcome and human override most coding agents send every fork to the frontier model and ask you to approve the rest this 12-page guide moves the forks to a decision layer and keeps you only on what you cannot undo Send this PDF and the original Jev article to Claude Code or Codex and start rebuilding one expensive agent decision at a time ↓

这条是引用着另一条写的：x.com/hanakoxbt/status/2101709924828934222