WHAT NEW IN X10.29 THU

X 上的 AI 合集档案 · 22 个合集 · 58,925 条

jevVIDEO

把 Sonnet/Haiku/Opus 分工写进 settings.json 一次配好

在 ~/.claude/settings.json 里固定主模型 Sonnet 5.5、子 agent 用 Haiku 5.5、Opus 5.5 只当顾问,免去每天敲四个 flag;帖尾称 Jev 在毫秒级接管无需思考的微分支。

Skunzu
@skunzu3 天前收录原推发布于 2026.10.25
Skunzu @skunzu原推发布于 2026.10.25
在 X 上打开原推

STOP TYPING THE SAME 4 FLAGS EVERY F*CKING MORNING Most people set up the Sonnet / Haiku / Opus split once, in a tweet, and lose it the next time they open a terminal. The fix: put the whole split in one file Claude Code reads on every launch. Same savings as the advisor setup, zero flags, and it never "forgets" which model to use. You just open ~/.claude/settings.json and paste this: { "model": "sonnet", "advisorModel": "opus", "effortLevel": "medium", "env": { "CLAUDE_CODE_SUBAGENT_MODEL": "haiku" } } then launch it like this: claude here's what that one file does: → Sonnet 5.5 runs every session and writes the code → every subagent (search, grep, docs) spins up on Haiku 5.5 → Opus 5.5 sits on call as advisor, it never types a line → medium effort by default, so routine turns stop overthinking Opus usually only gets called at three moments: → before a plan locks → when the same error shows up twice → before calling it done Set it once. Sonnet builds. Haiku digs. Opus judges. JEV engineering is the same move one layer down: the forks that need no thinker (which file, which model, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split the full tree > model: sonnet, the main session for every task > env CLAUDE_CODE_SUBAGENT_MODEL: haiku, every subagent by default > advisorModel: opus, reads the whole transcript only when called > effortLevel: medium, bump it per task with /effort when you need it > CLAUDE.md: when to call the advisor > JEV routes micro-forks in milliseconds paste this prompt into Claude Code ↓ "Make my model split permanent: 1. In ~/.claude/settings.json set model to sonnet, advisorModel to opus, effortLevel to medium, and add CLAUDE_CODE_SUBAGENT_MODEL=haiku under env. 2. Keep every other setting I already have. 3. Add one rule to CLAUDE.md: call the advisor before a plan locks, after the same error twice, and before calling a task done. 4. Route fast deterministic forks (tool choice, path selection, retries) through JEV. Show me the full diff first. No edits until I say go." ↳ http://code.claude.com/docs/en/settings save it, it's the last time you'll ever type --model

这条是引用着另一条写的:x.com/0xcodila/status/2109623431763292649

x.com/skunzu/status/2114071979296739614采集于 2026.10.25 10:04