Skip to content
Back to blog

· 3 min read

The Claude Code advisor "trick", checked

A viral post sold a Sonnet, Haiku and Opus setup in Claude Code as something Anthropic had just shared. I checked it against the docs and ran the command. Most of it is real; some of it isn't.

A post went viral this week presenting a Claude Code "goldmine" as something Anthropic had just shared: Sonnet runs the session, Haiku subagents search the repo, and Opus reviews as an advisor. I checked it against the official docs and ran the command. Most of it holds up; some of it doesn't.

What's real

The command works:

claude --model sonnet --advisor opus

The --advisor flag isn't listed in claude --help, and the docs say so. But inside the session, /advisor answers with the model acting as advisor.

The advisor has to be at least as capable as the main model. Try --advisor haiku on Sonnet and Claude Code refuses:

"haiku" cannot advise "claude-sonnet-5-5" (the advisor must be at least as capable as the main model). The advisor will not be used for the main model.

Haiku subagents exist too: you set model: haiku in a subagent's frontmatter. The built-in Explore subagent, though, runs on your session's model, not Haiku.

What's off

  • It isn't new. Anthropic published "The advisor strategy" on April 9, and the advisor has been in the Claude Code changelog since version 2.1.117, in April.
  • Opus doesn't step in at "three fixed moments". According to the docs, Claude tends to consult before committing to an approach, when an error keeps recurring and before declaring a task done, but the timing is "model-driven rather than rule-based".
  • The audit subagent may not see your changes. With isolation: worktree, its copy of the repo is branched from your default branch, not from your session's HEAD (subagent docs). Unless you change that, its copy doesn't have the diff it's meant to check.
  • "JEV" isn't Anthropic. It's Jev, a product from TypeSafe, added to the prompt as step five.
  • Cheaper than Opus all session, not free. Every advisor call processes the whole conversation, uncached. On startup Claude Code itself warns it "may use more tokens". The docs say pairing a faster main model with a stronger advisor typically costs less than running the stronger model throughout.

Two more details

  • The advisor only works with the Anthropic API: not on Bedrock, Vertex or Foundry.
  • It's turned off if you have DISABLE_TELEMETRY set.

What I tested and what I didn't

I ran the command, /advisor and the Haiku refusal. The consultation timing, the worktree base and the cost come from the docs: I haven't measured a real task with the advisor.

Have you tried it? Which pairing are you running?