Claude Code’s /advisor: Opus Guidance Without Opus for Every Task

0:00 Right, so you've got your coffee, your terminal's open, and Claude's about to chew through a big refactor. This is The Claude Code Changelog, brought to you by Jellypod AI. I'm Lachlan. And I'm James. Today is the advisor tool, the /advisor command. It's experimental, and it's the cleanest answer yet to a question we've circled before: how do you get Opus level thinking without paying Opus prices for every grep?

0:28 Yeah, this builds on the cost themes from a couple of earlier episodes. Pinning Haiku for the auto mode safety checks, and reining in runaway Opus threads. The advisor is the hybrid version of both ideas, for your everyday CLI work. Here's what it is. A cheaper executor model, say Sonnet 5.5 or Haiku 5.5, does the actual work.

0:52 When it hits something tricky, it consults a stronger model, like Opus 5.5 or Fable. The advisor sees the full conversation, every tool call and result, and sends back guidance. The advisor never calls tools and never writes to your transcript. It only steers. Now, what changed recently. Version 2.1.300 added /advisor on Claude Platform on AWS, and added CLAUDE_CODE_DISABLE_ADVISOR_TOOL set to 1, which removes the tool entirely and ignores any advisor settings you've configured.

1:27 Earlier point releases, 2.1.293 and 2.1.287, brought Haiku 5.5 and Sonnet 5.5 into the accepted advisor lineup. Setup has four routes. Type /advisor on its own and you get an interactive model picker. Type /advisor opus and it saves to your user settings across sessions. /advisor off turns it off. For a persistent default, put advisorModel set to opus in your global or project settings file.

1:57 And for a single session, pass the advisor flag at launch with opus after it. Quick note, per the 2.1.300 notes: no advisor runs until you pick one. It's opt in. Why bother? Two comparisons. First, opusplan. There, Opus plans upfront, then the session drops to Sonnet and Opus never speaks again. The advisor stays available the whole way through.

2:20 Second, subagents. Those run isolated tasks in their own context. The advisor reads your actual shared context, so its advice is grounded in what's already happened. On numbers, a video walkthrough we looked at reports that Sonnet with an Opus advisor scores roughly three percent higher on SWE bench than Opus alone, with token costs about twelve percent lower.

2:43 That's the reported figure, so treat it as a benchmark, not a promise for your repo. The logic is that file reads, searches, and boilerplate edits get billed at executor prices. As for when Claude calls the advisor, that's model driven, not rule based. It tends to consult on ambiguous architecture tradeoffs, on test or build failures that keep recurring, and before declaring a complex task done.

3:09 But you can force it. Put something in your prompt like: consult the advisor before finalizing the database migration schema.

3:16Or check with the advisor before rewriting the auth middleware.

3:19 You'll see an Advising row in the terminal while it's thinking. Afterwards it reads Reviewed, Declined with a reason, or Unavailable with an error code. Press Ctrl+O on that row and you can expand it to read exactly what the advisor said. Worth doing the first few times, just to calibrate how useful it is.

3:39 Caveats, and there are a few. The advisor has to rank at or above your main model. Haiku 4.5 can't be an advisor, and Sonnet can't advise Opus. If you configure a bad pairing, or your organization's availableModels allowlist excludes it, Claude Code either warns at startup or quietly runs without an advisor.

3:59 So if you think it's on and nothing's happening, check that. Provider support matters too. It needs the first party Anthropic API or Claude Platform on AWS. Amazon Bedrock, Google Vertex AI, and Microsoft Foundry aren't supported. One more, on caching. Changing /advisor in the middle of a session preserves your main conversation's prompt cache prefix, unlike /model, which invalidates it.

4:25 But the advisor's own read of the transcript isn't cached between calls, so every consultation reprocesses the history. Long sessions with frequent consults add up. And Claude treats advice as advice. If your files or command output contradict it, it surfaces the conflict instead of blindly following.

4:44 A few smaller wins from 2.1.301 to 2.1.303. Vim mode got the percent motion to jump to a matching bracket, alone or with d, c, or y. In 2.1.301, switching to a model with a smaller context window now has the previous model summarize first. In 2.1.302, the Read tool refuses image sources over 128 megabytes instead of loading them into memory, there's an Idle compaction row in /config, and hooks are capped at twelve running at once per event.

5:22 2.1.303 is general reliability fixes.

5:27So your homework run /advisor opus on a Sonnet session, give it a messy migration,

5:32 and hit Ctrl+O on the advice. See you next time.