Claude Code 分层架构:Opus 规划、Sonnet 执行、Jev 路由小决策
把文件选择、工具选择、重试逻辑等小决策交给 Jev 轻量决策层,避免用 Opus 反复思考。
claude code tip that can save you a f*cking crazy amount of opus tokens: stop making your most expensive model handle every single task. give it a team the architecture is simple: opus leads, sonnet executes, fable reviews division of labor > Opus 5.5 on high - breaks down the problem, makes the plan and handles the hard coding > Sonnet 5.5 on medium - explorer maps the codebase, worker implements changes and tests, researcher fetches documentation > Fable 5.1 through /advisor fable - acts as an independent reviewer instead of interrupting every step when to call the advisor -> before committing to a complex implementation -> after hitting the same error repeatedly -> when the agent claims a long task is finished Jev engineering pushes this further: route tiny decisions like file selection, tool choice and retry logic through a lightweight decision layer, instead of paying Opus to reconsider them on every iteration give Claude Code this prompt ↓ "Audit and upgrade my Claude Code agent architecture: 1. Inspect ~/.claude/agents and .claude/agents for existing explorer, worker and researcher agents. Reuse compatible ones and propose definitions only for roles that are missing. Configure new agents for Sonnet with medium effort. Flag any existing agents that explicitly require another model, without modifying them. 2. Review ~/.claude/settings.json and propose setting the main session to high effort and the advisor model to Fable. 3. Check for environment variables or configuration that may disable advisor functionality, including CLAUDE_CODE_DISABLE_ADVISOR_TOOL, DISABLE_TELEMETRY, feature-flag fetching issues and CLAUDE_CODE_EFFORT_LEVEL Report findings only. 4. Add a proposed rule to ~/.claude/CLAUDE.md to request advisor review before major plans, recurring-error investigations and final completion checks. Show all proposed changes as diffs. Do not edit any files until I approve."
这条是引用着另一条写的:x.com/l1vsun/status/2105630781578125651