← sources

What should I put in my Claude.md to make efficient use of Sub agents?

source: reddit (reddit:ClaudeAI)published 2026-10-22 16:44ingested 2026-10-23 02:38

retained text

I am trying to use sub agents to apply lower models to more simple tasks. Gemini Argon told me to put the following in Claude.MD

\>## Subagent Model Routing Rules

Your main session runs on Opus. Unless the user explicitly asks you to work directly in the main thread without subagents, dynamically route subagents using the \`Agent\` tool's \`model\` parameter:

1. \*\*Use \`model: "haiku"\` for read-only exploration & logs:\*\*

   \- Codebase searches (\`Explore\`), finding symbol references, mapping directories, and summarizing build/test/crash logs.

   \- Never use \`haiku\` subagents to edit code.

2. \*\*Use \`model: "sonnet"\` for isolated coding & test writing:\*\*

   \- Implementing well-scoped helper functions, writing unit tests, refactoring single modules, or running read-only code review checks.

3. \*\*Keep \`opus\` (main thread or \`Plan\` subagent) for complex architecture:\*\*

   \- Multi-file architectural design, subtle bug diagnosis, and final integration.

So far I don't see much of a difference in usage. What are you using in your Claude.md to accomplish this?

Edit: it seems to actually be worse than with nothing in Claude.md

document fetched 2026-10-23 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xcnrg6/what_should_i_put_in_my_claudemd_to_make