Recipe: Research with Multiple Perspectives
This recipe shows how to use multiple interlocutors to explore a topic from different angles. One interlocutor does research, another provides critique, and you can quickly get second opinions without derailing the main conversation.
The Setup
---
interlocutors:
- name: Researcher
prompt: |
You are a thorough researcher. When exploring a topic:
- Consider multiple sources and viewpoints
- Note uncertainties and limitations
- Suggest follow-up questions
provider: anthropic
model: claude-sonnet-4-20250514
tools:
- native: search
- think_about: >
What are the key questions here? What might I be missing?
What assumptions am I making?
- name: Critic
prompt: |
You are a skeptical critic. Your job is to:
- Challenge assumptions and weak arguments
- Point out missing evidence or alternative explanations
- Steelman opposing viewpoints
Be constructive but rigorous.
provider: anthropic
model: claude-sonnet-4-20250514
- name: Synthesizer
prompt: |
You synthesize discussions into clear summaries. Focus on:
- Key points of agreement and disagreement
- Open questions that remain
- Actionable conclusions
provider: anthropic
model: claude-3-haiku-20240307
---Usage
Start with research
:ask[Researcher] I'm trying to understand the tradeoffs between
microservices and monolithic architectures for a team of 5 developers
building a B2B SaaS product.The Researcher will explore the topic, potentially using web search and their thinking tool.
Get a quick critique
Use :aside to get feedback without switching the main conversation:
:aside[Critic] What's wrong with this analysis?The Critic responds, then the next message goes back to the Researcher automatically.
Permanently switch for deeper critique
:ask[Critic] Let's dig into the claim about "complexity tax." What's
the actual evidence here?Now you’re in a conversation with the Critic until you switch again.
Synthesize at the end
:ask[Synthesizer] Summarize this discussion. What did we learn? What
should we do next?Variations
Domain-specific experts
interlocutors:
- name: Legal
prompt: You are a legal expert. Focus on regulatory compliance...
- name: Technical
prompt: You are a senior engineer. Focus on implementation...
- name: Business
prompt: You are a business strategist. Focus on market fit...Agent delegation
Have one interlocutor call another as a tool:
interlocutors:
- name: Lead
prompt: You coordinate research. Delegate to specialists.
tools:
- agent: Researcher
name: research
usage: Get detailed research on a specific topic.
- agent: Critic
name: critique
usage: Get critical analysis of a claim or argument.
- name: Researcher
prompt: ...
tools:
- native: search
- name: Critic
prompt: ...Now the Lead can autonomously decide when to delegate:
:ask[Lead] Evaluate whether we should migrate from PostgreSQL to
CockroachDB for our multi-region deployment.Different models for different roles
Use cheaper/faster models for quick checks:
interlocutors:
- name: Deep
model: claude-sonnet-4-20250514 # For complex analysis
- name: Quick
model: claude-3-haiku-20240307 # For quick sanity checksTips
Use
:asideliberally. It’s cheap to get a second opinion without losing your place in the main conversation.Give each interlocutor a distinct voice. The prompts should produce noticeably different responses, otherwise there’s no point in having multiple speakers.
Use
:reset[]when switching topics. If you’re starting a new line of inquiry, reset context to avoid confusion from earlier discussion.