Claude Code 多模型分工：Opus 写脚本、Haiku 扫描、Jev 路由微决策

Belial @0xbelial原推发布于 2026.10.25

在 X 上打开原推

THIS IS A REAL F*CKING GOLDMINE Anthropic just shared one of the smartest tricks for Claude Code The idea: let Opus 5.5 turn your goal into a workflow script, then swarm the new Haiku 5.5 across the repo in parallel Massive sweeps with almost zero Opus burn: the script runs the loop and Opus only reads what survives You just launch Claude Code like this: claude --model opus --effort ultracode here's the split: → Opus 5.5 compiles your objective into a runnable workflow.js → Haiku 5.5 fans out across every file at 340 tok/s and flags candidates → Sonnet 5.5 patches each fix in its own isolated worktree and cross-checks the rest → you stop paying Opus to read dozens of raw agent transcripts Opus only shows up at three moments: → writing the workflow script before the run → reading the verified findings after the swarm → signing off before a single line hits git Opus scripts · Haiku sweeps · Sonnet patches JEV engineering is the same move one layer down: the forks that need no thinker (which file, which model, retry or stop) clear through JEV in milliseconds, and the big models only see the ones that actually split the full tree > Opus 5.5 runs the main session and authors workflow.js > pipeline() fans out 6 Haiku 5.5 auditors across every file > parallel() spins up 4 Sonnet 5.5 workers, one worktree each > verify runs lint, regression and rule checks on every diff > false positives get dropped before Opus ever sees them > Opus 5.5 reads only the findings that survived > JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the main session on Opus 5.5 with ultracode effort 2. In every workflow, put read-only scan stages on Haiku 5.5 and edit stages on Sonnet 5.5, each edit in its own isolated worktree 3. Add one rule to CLAUDE.md: every finding passes an independent verify agent before it's reported, and fix loops stop after two rounds with no progress 4. Return verified findings only to the main session, never raw agent output 5. Route fast deterministic forks (tool choice, path selection, retries) through JEV Show me every change as a diff first, no edits until I say go" ↳ http://code.claude.com/docs/en/workflows save it, this one pays for itself on the first big refactor

这条是引用着另一条写的：x.com/0xtrackmind/status/2113644871529418857