Claude Code 分层：Opus/Sonnet/Fable 加 Jev 决策

kocer @kocer_eth原推发布于 2026.10.03

在 X 上打开原推

this is f*cking insane. this tip saved me thousands of dollars. finding a file, checking documentation, and deciding your architecture do not need the same model. the useful part of this setup is knowing where to spend the expensive reasoning. Opus 5.5 › planning and hard decisions. keep it on architecture, ambiguous bugs, and reviewing changes that could break something important. Sonnet 5.5 › scoped execution. give it 3 specific jobs: explore the code, make the approved edit, or research the documentation. set these workers to medium effort instead of letting every subtask inherit your main model. Fable 5.1 › a second opinion. use `/advisor fable` if your account supports it. request a consultation before a risky approach, after a recurring failure, or before accepting complex work. it receives the full conversation when called. this adds usage, so don't mistake it for free background review. JEV › narrow, structured decisions. in a custom integration, let it choose between predefined options: relevant notes, worker tiers, or recovery paths. it returns structured results, not another essay. it is not a built-in Claude Code setting. say you're fixing a checkout bug: Sonnet locates the failing path and reproduces it. Opus decides whether the fix needs a design change. Sonnet applies the approved patch and runs the relevant tests. request Fable's opinion if the failure repeats or the result remains uncertain. paste this into Claude Code to prepare the configuration. it covers Opus, Sonnet, and Fable. connecting JEV requires a separate integration. "prepare my Claude Code setup for this division of work. inspect first. do not change files or settings until i approve the proposed diffs. check ~/.claude/agents and .claude/agents for existing explorer, worker, and researcher roles. reuse matching agents. only propose new definitions for missing roles. explorer locates relevant code without editing. worker makes scoped changes and runs relevant tests. researcher checks documentation and returns sources. for these workers, propose model: sonnet and effort: medium. if an existing agent explicitly uses another model, preserve it and report the difference. check which model versions the aliases resolve to. propose Opus 5.5 as the main model for planning, difficult decisions, and final review. configure high effort specifically for that model using the supported per-model setting. do not assume a legacy global effortLevel setting applies to Opus 5.5. check whether Fable advisor is supported by my account, Claude Code version, and connection. if available, propose advisorModel: fable in ~/.claude/settings.json. explain any usage-credit consent required. do not accept billing changes for me. check CLAUDE_CODE_DISABLE_ADVISOR_TOOL, DISABLE_TELEMETRY, and anything blocking feature-flag fetching. also inspect CLAUDE_CODE_EFFORT_LEVEL and other model or effort overrides. report conflicts without changing environment variables, privacy controls, or permissions. propose this instruction for ~/.claude/CLAUDE.md: consult the advisor before a risky plan and after the same error occurs twice. also consult it before accepting complex work as complete. these are consultation instructions, not guaranteed automatic triggers. preserve unrelated settings and existing instructions. show every proposed change as a diff. list anything requiring manual action, then wait for my approval." compare usage on similar tasks. cheaper workers only help if retries and extra consultations don't eat the difference. spend on the decision that needs intelligence, not every step around it.

这条是引用着另一条写的：x.com/kocer_eth/status/2106078938397286811