WHAT NEW IN X10.30 FRI

X 上的 AI 合集档案 · 22 个合集 · 61,986 条

jevVIDEO

Claude Code 按任务调 effort,简单路由交给 Jev

用 Opus 低 effort 跑主会话、按需切 xhigh,子 agent 各自设 effort,文件选择/工具选择/重试这类简单分叉交给 Jev 毫秒级处理。

Livsun
@L1vsun12 天前收录原推发布于 2026.10.17
Livsun @L1vsun原推发布于 2026.10.17
在 X 上打开原推

THIS IS F*CKING BRILLIANT you're probably wasting Opus tokens on dumbass tasks that never needed that much reasoning in the first place Claude Code lets you control effort levels. most people just leave the dial alone and pay for the same level of thinking on a fucking rename as they do on a complex architectural decision start with: claude --model opus --effort low here's the move: → low for routine edits, imports, renames and formatting → /effort xhigh when a problem actually needs deep reasoning → give subagents their own effort settings through frontmatter → stop treating every tiny task like a goddamn research project but here's where you turn the dial up: → before a complex plan touching multiple files → when the same bug survives repeated fixes → before a high-impact change that's painful to reverse low for the boring shit. xhigh when the stakes justify it. JEV engineering applies the same principle one layer down: route simple forks like file selection, tool choice and retries through a lightweight decision system, instead of wasting expensive model reasoning on every tiny fucking decision the full tree > Opus runs the main session at low effort > switch to xhigh when deeper reasoning is needed, then dial it back down > search uses Haiku at low effort > review uses Sonnet at high effort > each subagent gets an effort level suited to its role > JEV handles simple routing decisions in milliseconds paste the tree and this prompt into Claude Code ↓ "Review my Claude Code setup and propose changes based on this structure: 1. Configure the main session to use Opus at low effort. 2. Set low effort for read-only subagents and high effort for the review subagent, where supported. 3. Add a CLAUDE.md rule to increase effort for complex multi-file plans, repeated unresolved errors, and high-impact changes, then return to low effort afterward. 4. Identify simple deterministic decisions that could be routed through JEV. Show me every proposed change as a diff first. Do not edit anything until I approve." ↳ https://code.claude.com/docs/en/model-config save this shit before you forget it

这条是引用着另一条写的:x.com/edgerunpro/status/2110730246014025852

x.com/l1vsun/status/2111449660653183041采集于 2026.10.18 10:04