Claude Clue #33: Let Your AI Phone a Smarter Friend

Remember "Phone a Friend" in Who Wants to be a Millionaire? It's like that, but with lower stakes.

Claude Clue #33 - Tips for working with AI
Christopher's Claude Clues #33

Claude Code has an experimental feature called the advisor. You pair your daily-driver model with a stronger one, and Claude consults it at the moments that decide how the whole task goes. It tends to call before committing to a plan, when the same error keeps coming back, and before declaring the job done. The advisor reads the full conversation and returns guidance that your main model applies before continuing. Run /advisor opus in a session and you're set. Claude decides when to dial [the friend].

Roscoe runs it the way Anthropic intended, with Sonnet doing the work and Opus a phone call away for planning and troubleshooting. This seems to make a big difference. Anthropic benchmarked the pair (Sonnet with an Opus advisor), and it scored higher on SWE-bench than Sonnet alone while costing 12% less per task.

For now, this lives in the terminal and the API only; Cowork and Claude Desktop don't have it, which makes it a real difference between the two doors from Clue #32. And the advisor must be at least as capable as your main model, so the call always goes up, never down. Right now, though, Fable isn't available as an advisor. I find this a little strange, since this pattern of advising or planning is used by other systems too, and I would expect Fable to be a great advisor.

One Claude drafted this. Another was on call, just in case.

Thanks to Tim "Roscoe" Rawson, whose daily driver has Opus on speed dial.

Subscribe to Almost Entirely Human

Don’t miss out on the latest issues. Sign up now to get access to the library of members-only issues.
jamie@example.com
Subscribe