diverge

Diverge: Multi-Perspective Divergent Research

Spawn N independent agents with uncorrelated context windows to explore a question from different angles, then auto-synthesize their findings into a unified analysis and PRD.

Arguments

[N] "research question or topic"

• N is the number of agents (default 3, min 2, max 7).
• The question is the research topic. Missing or unclear: ask the user to clarify before starting.

Phase 1: Frame the Research

Before spawning agents, frame the research space. Write a short research brief (3-5 bullet points) that:

• States the core question
• Lists known constraints or context from the conversation
• Identifies 2-3 dimensions of exploration (e.g., technical feasibility, workflow design, prior art)

Present this to the user and confirm before proceeding. Adjust if the user gives feedback.

Phase 2: Spawn Independent Agents

Launch all N agents in parallel using the Agent tool. Each agent MUST:

1. Have a unique research lens: assign each a distinct angle, perspective, or methodology. Examples:

   • "Prior art and industry patterns" (what exists, what others do)
   • "First-principles technical design" (bottom-up from constraints)
   • "User experience and workflow" (developer ergonomics, day-in-the-life)
   • "Failure modes and risks" (what can go wrong, edge cases)
   • "Scale and evolution" (how this grows, maintenance burden)
   • "Contrarian/devil's advocate" (challenge assumptions, explore alternatives)

2. Receive the same research brief but with their unique lens clearly stated

3. Be instructed to:

   • Research independently (web search, codebase exploration, reasoning)
   • Produce a structured output with: Key Findings (3-5), Concrete Recommendations (2-3), Open Questions, and a Confidence Assessment
   • NOT be told what other agents are exploring
   • Think creatively within their lens, surprising or non-obvious insights are more valuable than safe ones

4. Use subagent_type: "general-purpose" (they need web search, file access, etc.)