Diverge

Multi-perspective divergent research. Spawns N independent agents with uncorrelated context windows to explore a question from different angles, then auto-synthesizes findings into a unified analysis and PRD.

Arguments

$ARGUMENTS — format: [N] "research question or topic" where N is optional (default: 3)

Parse Arguments

Extract:

• agent_count: the optional leading integer (default 3, min 2, max 7)
• research_prompt: the quoted or unquoted research question/topic

If the research prompt is missing or unclear, ask the user to clarify before proceeding.

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)

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

1. 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

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

1. Do NOT use worktrees (research only, no code changes)

Agent prompt template (customize the lens per agent):

You are a research agent exploring a question from a specific perspective.

## Research Brief
{research_brief}

## Your Lens: {lens_name}
{lens_description}

Explore this question deeply from your assigned perspective. Use web search, read relevant files in the codebase, and reason carefully.

## Output Format
Return your findings in this exact structure:

### Perspective: {lens_name}

#### Key Findings
1. [Finding with supporting evidence or reasoning]
2. ...
(3-5 findings)

#### Concrete Recommendations
1. [Actionable recommendation with rationale]
2. ...
(2-3 recommendations)

#### Open Questions
- [Questions your research surfaced but couldn't resolve]

#### Confidence Assessment
- Overall confidence: [High/Medium/Low]
- What would increase confidence: [specific next steps]

#### Surprise Factor
- [1-2 insights that were unexpected or counterintuitive]

Phase 3: Synthesize

After ALL agents return, produce a unified analysis:

Synthesis Structure

1. Convergence Points Where did multiple agents independently arrive at similar conclusions? These are high-confidence signals.

2. Divergence Points Where did agents disagree or propose conflicting approaches? Highlight the tension and what drives it.

3. Unique Insights Novel findings that came from only one agent's lens but are valuable. Flag the "surprise factor" items.

4. Consolidated Recommendations Merge and prioritize all recommendations. For each:

• State the recommendation
• Note which agents support it (and from what angle)
• Flag any dissenting views
• Assign priority: must-have / should-have / nice-to-have

5. Open Questions & Risks Aggregated unknowns that need resolution before moving to implementation.

6. Per-Agent Highlights For each agent, include a 2-3 sentence summary of their most distinctive contribution.

Phase 4: Draft PRD

Based on the synthesis, draft a mini-PRD markdown file:

# PRD: {topic}

## Problem Statement

{1-2 paragraphs from synthesis}

## Goals & Non-Goals

### Goals

- ...

### Non-Goals

- ...

## Requirements

Each requirement MUST include verifiable acceptance criteria — a concrete condition that can be checked by running a command, reading output, or inspecting behavior. "Works correctly" is not verifiable. "Returns 200 with JSON body containing 'id' field" is.

### Must-Have

- Requirement: ...
  - Acceptance: {specific, testable condition}
- ...

### Should-Have

- Requirement: ...
  - Acceptance: {specific, testable condition}
- ...

### Nice-to-Have

- Requirement: ...
  - Acceptance: {specific, testable condition}
- ...

## Design Considerations

{Key tensions and trade-offs from divergence points}

## Open Questions

{From synthesis}

## Research Provenance

{Which lenses contributed, key convergence/divergence summary}

Save the PRD to the working directory as prd_{slugified_topic}.md.

Present the full synthesis AND the PRD path to the user. Ask if they want to refine the PRD before using it with /diverge-prototype.

Rules

• Independence is critical: agents must NOT share context. Do not include other agents' findings in any agent's prompt.
• All agents launch in a single parallel batch: use one message with N Agent tool calls.
• No filtering: include all agent findings in the synthesis, even contradictory ones. Tension is signal.
• Attribute insights: always note which lens produced which finding.
• Be honest about confidence: if synthesis is weak in an area, say so.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sjarmak
• Source: sjarmak/agent-workflows
• License: MIT

Install and usage instructions live in the source repository linked above.