Claude Code three-model split: Sonnet runs, Haiku scouts, Opus advises
A Claude Code setup where Sonnet implements, read-only Haiku subagents map the repo and docs, and Opus is consulted at three checkpoints; Jev is suggested for deterministic routing if already installed.
this claude code setup is f*cking gold stop paying opus to do sonnet's job let sonnet run the session, give haiku the discovery work, and keep opus available for the decisions that deserve a second opinion start with: claude --model sonnet --advisor opus here's the split to configure: → sonnet owns implementation and edits the code → haiku searches the repo and pulls relevant docs → opus advises on the approach, recurring failures and completion the advisor gets the full conversation when consulted, so you don't have to reconstruct the entire debugging session for it the three checkpoints i'd ask for: 1. before committing to a substantial plan 2. after the same error returns twice 3. before declaring a long task finished sonnet keeps building haiku brings back evidence opus challenges the important decisions the discovery agents should have narrow jobs: search → map the relevant files and modules grep → find callers, references and existing patterns docs → check the actual API documentation all read-only no three agents “helpfully” editing the same file while you're trying to understand it if JEV is already wired into your stack, keep it on deterministic routing where it helps. it isn't required to use the advisor paste this into claude code: “Propose this setup for my existing configuration: 1. Sonnet as the main model, with Opus as advisor. 2. Three read-only Haiku subagents: search for repository mapping, grep for call sites, and docs for primary API references. 3. Add instructions to consult the advisor before substantial plans, after the same error twice, and before completing long tasks. 4. Inspect existing agents first. Reuse suitable ones and flag conflicting settings. 5. If JEV is installed, show where deterministic routing could use it. Otherwise leave it out. Show every proposed change as a diff. Do not edit anything until I approve.” one catch: advisor calls still consume tokens, and consultation timing is model-driven rather than a hard three-call limit use it where a bad decision would waste more work than the review costs save this for your next refactor
This post quotes another post: x.com/sector_sol/status/2093769463065702779