uygnoey Overview
This skill captures the “advisor strategy” described in the Claude blog post and provides an implementation template for using the Claude Platform advisor tool in a Messages API request.
When to use
Use this approach when:
- Your agent can usually make progress with a cost-effective model (executor) using tools and iteration.
- Occasionally, the agent reaches a decision point where higher-quality reasoning would materially improve outcomes.
- You want to cap and measure the higher-cost model’s usage (e.g., max advisor consultations per request).
Instructions
- Choose an executor model (e.g., Sonnet or Haiku) to run the task end-to-end.
- Choose an advisor model (e.g., Opus) to provide guidance only when invoked.
- Add the advisor tool entry to your Messages API
toolslist and set:model: the advisor modelmax_uses: maximum number of advisor consultations per request
- Keep the advisor’s role limited to guidance:
- It should return a plan, correction, or stop signal.
- It should not call tools.
- It should not produce user-facing final output.
- Track costs and behavior:
- Review the usage block for separate advisor token accounting.
- Evaluate executor-only vs executor+advisor vs advisor-only to validate impact.
Examples
- Messages API example (from the post): examples/messages_api_example.py