Experimenting with settings to get more work done

End of the year is going to be busy with closing out projects and expanding what I do with Opus 5.5. I have not done a lot with tuning my token use within my account. The first time was when Fable 5 came out. It was token hungry, so I noticed the pattern that subagents tend to be the same model that I selected for my main conversation in Claude Code. This was when I found out on X that you can use CLAUDE_CODE_SUBAGENT_MODEL to set your subagent model within the env block of your settings.json file. While I was using Fable, I set my subagents to Opus 5, and it helped slow down the token burn. The subagents being created were also Fable. So I figured it is time to tune my settings to give me a decent spread of how models are used within the harness.

In discussions with my AI group, there was concern about losing Fable intelligence by switching back to Opus 5.5. There is a command that you can use called /advisor (docs still say it is experimental). What this does is set Claude Code to escalate hard decisions with the advisor tool. The examples given are below.

1. Before settling on an approach
2. When it is stuck on a recurring error
3. Before calling a task finished

I personally will not enable this until Fable 5.5 is out since I am happy with Opus 5.5 intelligence, but this is a good way to escalate when running loops or to get a second opinion. The other settings I'm putting in place for token efficiency:

1. CLAUDE_CODE_SUBAGENT_MODEL=sonnet for general subtasks
2. A Haiku subagent limited to Glob, Grep and Read. Claude hands it "where is X across the repo" searches, not understanding what is in a file.
3. In my CLAUDE.md global file I have set the caveat that all code creation will be done with Opus, to steer code writing away from Sonnet.

Claude created an agent file in ~/.claude/agents/ that pins Haiku to use Glob, Grep and Read commands. If you bring it up with Claude it will make the changes for you. Your mileage may vary with these settings but I figured I'd share in case you are like me and have not tried to tune your environment to be more efficient. Good luck to all. Make something interesting, and if you have better suggestions, please share.