Wink Pings

Don't let Opus write all the code: Claude Code's advisor play

Anthropic added an advisor mode to Claude Code: Sonnet executes, Haiku explores, and Opus steps in only for planning, repeated errors, and before completion. Includes configuration commands and a demo video.

Anthropic has added an advisor tool to Claude Code. It lets a stronger model act as a consultant at key moments in a task, instead of doing everything from start to finish.

X user @xsmryze calls this trick 'real gold' and also provides a concrete configuration. First, look at the official demo:

Claude Code is Anthropic's command-line coding assistant, supporting models with different capabilities: Opus is the strongest and most expensive, Sonnet is balanced, and Haiku is the fastest and cheapest. Most everyday coding does not need Opus to take part, but solution selection, difficult errors, and pre-delivery checks need its judgment. Advisor mode is meant to resolve this contradiction.

Startup is simple:

claude --model sonnet --advisor opus

Sonnet 5.5 runs the session and writes code as the main model. Haiku 5.5 explores the repository and documentation in parallel, and can be split into three read-only subagents: search, grep, and docs. Opus 5.5 acts as the advisor, reading the full conversation at key moments and giving suggestions. There is also JEV responsible for millisecond-level simple branch routing, such as choosing files, choosing tools, and deciding whether to retry or stop; the large model only handles forks that truly require thinking.

Opus steps in only at three moments:

  • Before the plan is locked
  • When the same error occurs twice
  • Before a task is declared complete

These three things are exactly what determine a task's direction, blockers, and quality. At other times, Sonnet does the work, Haiku runs errands, and Opus only makes judgments.

Official documentation confirms that the advisor receives the full session, including every tool call and result, and then returns guidance. Claude decides when to consult, generally the three scenarios above. On cost, advisor calls are billed at its token price, but because they are triggered only at decision points, the total cost is usually lower than running Opus throughout. And enabling or disabling the advisor does not break the main model's prompt cache.

@xsmryze also provides a template that can be pasted directly into Claude Code:

Rebuild my Claude Code setup around this tree:
1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor.
2. Create three read-only Haiku 5.5 subagents: search (maps the repo), grep (finds call sites), docs (pulls the API reference).
3. Add one rule to CLAUDE.md: consult the advisor before locking a plan, after the same error twice, and before calling a task done.
4. Route fast deterministic forks (tool choice, path selection, retries) through JEV.
Show me every change as a diff first. No edits until I say go.

This mode is like that senior engineer on a team who only appears at key reviews: they do not bother you normally, but can check things at important milestones. If you often run long tasks, this trick is worth saving.

Official documentation: https://code.claude.com/docs/en/advisor

发布时间: 2026-11-21 21:42