Opus Should Only Review Three Times: A Cost-Saving Multi-Model Workflow for Claude Code

Opus is very capable, but asking it to find files in the repo is like using a sledgehammer to crack a nut. In the tweet, xsmryze breaks down a multi-model division of labor for Claude Code: Sonnet does the main writing, Haiku searches and reviews in parallel, and Opus serves only as an advisor. The core goal is straightforward: don't waste the most expensive model and the main-session context on grunt work.

Claude Code is Anthropic's terminal coding agent; it can read and write files in the repo, run commands, and edit code. Once you add more models, a problem appears. Search output, grep results, and diff audits pile up in the main session, and the context quickly gets polluted. The solution the tweet author offers is to divide work according to what each model is good at:

• Sonnet 5.5 runs the main session, writes code, applies fixes, and coordinates tasks.
• Haiku 5.5 spins up three subagents to work in parallel: search maps the repository, grep finds call sites, and audit checks the diff in a temporary worktree.
• Opus 5.5 is not always on; it only acts as advisor. The tweet lists three intervention points: before the plan is locked in, when the same error appears a second time, and before declaring completion.
• The tweet author also mentions JEV, treating it as a lower-level fork router: choices such as which file to select, which tool to use, and whether to retry or stop do not need a large model's reasoning, so they are handled by JEV at millisecond speed; the large model only handles things that involve real disagreement.

One sentence summarizes the division of labor in the tweet: Sonnet builds. Haiku swarms. Opus reviews. In other words, Sonnet is responsible for building, Haiku for swarm searching, and Opus for reviewing.

The most valuable part of this setup is not the model combination but context isolation. Some commenters asked: don't Haiku and Sonnet still read the same repository? How do you save tokens? The key is that search results, grep call sites, and diff audits each run in their own context, and in the end only summaries and judgments are sent back to the main session. The main session is no longer stuffed with screen after screen of file contents, and Opus does not have to start its review by finding files. The audit is also placed in a throwaway worktree, so checks do not touch the working branch. Exactly how many tokens you save depends on the repository and the task, but the direction is clear.

The tweet gives a startup command: claude --model sonnet --advisor opus. Don't rush to copy it verbatim. Commenter Kuz reminded people in a reply that Opus advisor mode has existed since April, and in Claude Code it is /advisor; the Haiku subagent model is set with the environment variable CLAUDE_CODE_SUBAGENT_MODEL, not a command-line flag. Check the docs before configuring. Docs link: http://code.claude.com/docs/en/sub-agents

The tweet author also provides a copy-paste-ready configuration prompt, with five core steps. Organized here in English according to the original meaning:

Reconfigure my Claude Code around this division of labor:
1. Run the main session on Sonnet 5.5, with Opus 5.5 as advisor.
2. Create three Haiku 5.5 subagents in .claude/agents: search maps the repository, grep finds call sites, and audit checks the diff.
3. Set audit to isolation: worktree so checks do not touch the working branch.
4. Add a rule to CLAUDE.md: hand repository search and test runs to Haiku 5.5 subagents to keep the main context small.
5. Route fast deterministic forks such as tool choice, path choice, and retry decisions through JEV.
Show me the diff of all changes first; do not change anything until I say go.

The last step is quite important. Looking at the diff before letting the agent act can save a lot of rework. Don't start a three-model party directly on the main branch; try it on a branch. Small scripts and small repositories may not need this scheduling, while large repositories, long sessions, and search-heavy tasks benefit more clearly.

Original post: https://x.com/xsmryze/status/2117335920827371778