Sonnet works, Opus advises, Haiku runs errands: stretching your Claude Code quota

You do not need Opus for the whole session. Here is how I split work across three models in Claude Code, with the one-time setup, everyday use, how to check that it saves anything, and templates you can paste.

The easiest way to burn through your quota in Claude Code is to run the strongest model for the whole session. Most of what happens in a session does not need it: editing files, implementing an agreed plan, or finding where a function lives are all things Sonnet, or even Haiku, handles well. What really needs a strong model is a few moments: choosing a plan, getting unstuck, and one last check before calling it done.

This post is the usage guide I wrote for my team, turned into an article. The idea is that each of the three models does what it is good at, and a few lines in CLAUDE.md control when to call whom.

The idea: three models, each doing what it is good at

Day-to-day work goes to Sonnet. Opus is called in as an advisor only at key moments. Errands go to Haiku. The official documentation says that a faster main model paired with a stronger advisor is usually cheaper than running the strong model all the way through.

Role Model Responsible for When it runs
Main model Sonnet Writing code, editing files, everyday conversation The whole time
Advisor Opus Choosing a plan, debugging when stuck, a check before finishing Only when Claude decides it needs one
Errand sub-agent (scout) Haiku Finding files, finding functions, reading long documents or logs When it is given a task

One thing to know: every time the advisor is called, Opus rereads the entire conversation, and that part is not served from cache. So the point is not to ask more, it is to ask less and ask at the right moments. That is why the rules go into CLAUDE.md: there is no setting that limits how many times the advisor is called, so instructions are the only lever.

One-time setup (about 5 minutes)

Update Claude Code to the latest version first, then do the three steps below. Step 1 is for each person to do on their own. Steps 2 and 3 are best committed to the project repository so teammates get them on pull.

  1. Set the main model and the advisor (each person): open Claude Code in any project, type /model sonnet, then /advisor opus. The settings are saved to ~/.claude/settings.json and apply to every project. The desktop app accepts these two commands too.
  2. Add the “division of labor” rules to the project’s CLAUDE.md: paste the “division of labor” section from the templates at the end into the CLAUDE.md at the project root (create it if there is none).
  3. Add the scout sub-agent: put the scout.md from the end at .claude/agents/scout.md in the project (shared with the team). If you want it only for yourself and in every project, put it at ~/.claude/agents/scout.md.

Where each file goes:

File Shared with the team (committed to the repo) Yours only, all projects
CLAUDE.md <project>/CLAUDE.md ~/.claude/CLAUDE.md
scout sub-agent <project>/.claude/agents/scout.md ~/.claude/agents/scout.md
Model settings Not recommended (everyone’s quota differs) /model and /advisor write to ~/.claude/settings.json automatically

On macOS, ~ is /Users/<account>/. .claude is a hidden folder; in Finder, press Cmd + Shift + . to see it. If ~/.claude/settings.json already exists, do not overwrite it by hand. Use the commands instead.

You can also create a personal scout directly in Terminal (save the scout.md from the end to your Downloads folder first):

mkdir -p ~/.claude/agents
mv ~/Downloads/scout.md ~/.claude/agents/scout.md
ls ~/.claude/agents

Check that it took effect: /advisor should show Opus; /memory shows which CLAUDE.md files were read; ask Claude “list the sub-agents you can use” and you should see scout. Newer versions no longer have an /agents command.

Everyday use

Once it is set up, just work as usual. Claude follows CLAUDE.md to decide when to ask the advisor and when to send out the scout. Here is what to say when you want to step in yourself:

Situation What you can say
Something big, and you want a plan first “Ask the advisor about this approach before you start”
The same error will not go away “This error has come up twice now, ask the advisor first”
You want one more check before wrapping up “Have the advisor check it once before you say it is done”
Finding a file or where a function is “Use scout to find where the Prisma schema is”
Reading a long log or document “Use scout to read this log and report only the key errors”
Small edits (renames, wording, adding types) Just say it, no advisor needed

A few small tips:

  • When a conversation gets long, compress it with /compact or start over with /clear. The longer the conversation, the more it costs for the advisor to reread it each time.
  • Do one thing per conversation. Start a new one when the task changes.
  • When the task is short anyway, run /advisor off first to skip the advisor’s overhead.

Checking whether it saves anything, and how to adjust

After using it for a while, type /usage to see the usage by model. The advisor’s usage is counted in the totals there.

What you see What it means How to adjust
Mostly Sonnet, a small slice of Opus The setup is working Leave it alone
Opus usage is high The advisor is called too often Tighten the conditions in CLAUDE.md, for example “only ask when touching 5 or more files”
Almost no Opus The advisor is not active, or the tasks are too short Run /advisor to check the setting; it may also simply not be needed
Haiku is never used The scout is not being dispatched Check where scout.md is, or say “use scout to…” directly

How much you save depends on how you worked before. If you used Opus the whole time, you will save a lot. If you were already on Sonnet, the advisor is extra overhead, and what you get in return is fewer detours. The figures wmedia compiled from Anthropic’s tests are 11.9% lower cost and 2.7% higher accuracy than Sonnet alone. It depends on the task, so treat that as a reference, not a guarantee.

Limitations and common questions

  • The advisor is still an experimental feature: behavior, pricing and availability can all change.
  • It needs a direct connection to the Anthropic API: subscription accounts and API-billed accounts both work. Amazon Bedrock, Google Cloud and Microsoft Foundry are not supported.
  • Turning off telemetry also turns off the advisor: if you set an environment variable like DISABLE_TELEMETRY that turns off feature flags, the advisor cannot be enabled.
  • The advisor cannot be weaker than the main model: when the main model is Opus, the advisor can only be Opus or Fable; setting Sonnet simply will not attach.

Q: Can I set a maximum number of times the advisor is asked? No, there is no such setting in the official docs. You can only ask it, in CLAUDE.md or in the conversation, to ask more or less.

Q: How do I see what the advisor said? A line labelled Advising appears in the conversation; press Ctrl+O after it finishes to see the content.

Q: What if I do not want the advisor for now? Type /advisor off. To turn it back on, type /advisor opus.

Q: Do background sub-agents burn a lot of quota too? According to FourWeekMBA, versions after 11/04 switched to Haiku 5.5 to write status from short notes, which makes each update about 99% cheaper and a whole session about 5% lighter. Updating to the latest version applies it automatically.

How I use it myself

The shared CLAUDE.md for all my projects has the rules from the appendix below: the advisor is called at three moments only, and I handle everything else myself.

  1. Before choosing a plan: when a change touches 3 or more files, changes the architecture, or has several ways to do it, ask the advisor before starting.
  2. The second time the same error appears: if the same error or test failure survives two attempts in a row, stop and ask the advisor, rather than guess a third time.
  3. Before declaring it done: before saying “finished”, have the advisor check it.

I find the third one the most worthwhile. It is not about saving quota, it is about forcing one independent check before wrapping up. As for whether the threshold should be 3 files or 5, use the default for a while, then adjust by looking at /usage.

Appendix: templates

The “division of labor” section for CLAUDE.md

Paste it into the project’s CLAUDE.md. Adjust the thresholds to your team’s habits (for example “3 or more files”).

## Division of labor (saving quota)

### Call the advisor only at these three moments
1. Before choosing a plan: when a change touches 3 or more files, changes the architecture, or has several ways to do it, ask the advisor before starting.
2. The second time the same error appears: if the same error or test failure survives two attempts in a row, stop and ask the advisor.
3. Before declaring it done: before saying "finished", have the advisor check it.

Do not call the advisor for anything else: renames, wording changes, adding types, and implementing an agreed plan are all yours to handle.

### Errands go to the Haiku sub-agent
- Searching for files, finding where a function is, reading long documents or logs: hand it to the scout sub-agent.
- Ask it to return only conclusions and file paths. Do not bring whole files back into the main conversation.
- Work that needs judgment, design, or writing code stays with the main model.

scout.md

Put it at .claude/agents/scout.md (team) or ~/.claude/agents/scout.md (personal).

---
name: scout
description: A read-only sub-agent for errands. Use it to search for files, find where a function or setting lives, and read long documents or logs and summarize the key points. Do not use it when judgment, design or writing code is needed.
tools: Read, Grep, Glob
model: haiku
---

You are a fast, concise scout.

- Read only. Do not modify any file.
- Reply in this format:
  1. A one-sentence conclusion
  2. Relevant file paths and line numbers (path/to/file.ts:42)
  3. If needed, a key snippet of at most 10 lines
- Do not paste whole files, and do not suggest changes unless asked.
- If you cannot find it, say so directly and list where you searched.

References

Albert Shih

15+ years in software and DevOps, focused on AWS cloud architecture. Contact