Research Swarm

Phase 2: Check Prior Knowledge

Search for existing work on this topic before going external.

Step 1: Try Smart Connections MCP

mcp__smart-connections__semantic_search(
  query="[research topic]",
  limit=10,
  min_similarity=0.5
)

If Smart Connections is available, use results to inform angles.

Step 2: If Smart Connections unavailable, try Glob

1. Check CLAUDE.md for a FILE LOCATIONS table with vault/knowledge base path
2. Try common locations: ~/Documents/*/Zettelkasten/, ~/Obsidian/*/
3. If a vault is found, search with Grep for the topic keywords

Step 3: Evaluate results

• Strong match: Prior research exists — ask: "Found prior research on [topic] from [date]. Update it or start fresh?"
• Useful context: Related notes that inform angles — incorporate and proceed
• Nothing relevant: Proceed with full external research

Step 4: If no vault found

Skip this phase. Note in report: "No personal knowledge base detected — skipped prior research check."

Skip this phase if user says "fresh research" or "skip vault."

Phase 3: Decompose the Question

Break the topic into 3-5 independent research angles. Use mission context to tailor them.

With mission context loaded:

• Instead of generic "practical angle" → "How would [niche] coaches apply this?"
• Instead of generic "market angle" → "What does this mean for [client avatar]?"
• Instead of generic "integration angle" → "How would this fit a [delivery model] practice?"

Standard angle types:

• Technical: How does this work?
• Practical: How are people using this?
• Competitive: What alternatives exist?
• Integration: How would this fit your coaching practice?
• Cost/feasibility: What are the constraints?

Video research (optional): If my-skills/youtube-processor/ exists, include a video angle:

• "What are experts saying in recent videos about [topic]?"
• Sub-agent uses youtube-processor to extract transcripts and insights

If youtube-processor is not available, skip the video angle and use only web research.

Phase 4: Spawn Research Agents

Launch all angles in parallel using the Task tool:

For each angle:
  Task(
    description="Research [angle name]",
    prompt="[Specific research question for this angle]",
    subagent_type="general-purpose",
    model="haiku",
    run_in_background=true
  )

Launch ALL agents in a single message to maximize parallelism.

Cap at 5 sub-agents per swarm (diminishing returns beyond that).

Phase 5: Gather Results

Wait for agents to complete, then retrieve findings:

For each agent:
  TaskOutput(task_id=agentId, block=true)

Phase 6: Synthesize

Combine findings into a unified report. Look for:

• Agreements: Where do multiple angles converge?
• Contradictions: Where do they disagree? (This is where insight hides.)
• Surprises: What was unexpected?

Phase 7: Generate Report

# Research Report: [Topic]

**Research Date:** YYYY-MM-DD
**Angles Investigated:** [List]

## Executive Summary
[2-3 sentence synthesis of key findings]

## Prior Knowledge
[Links to related notes found in Phase 2, or "No prior research on this topic"]

## Findings by Angle

### [Angle 1]
[Key findings, with sources]

**Verdict:** [Act now | Bookmark | Ignore | Needs more research]
[One sentence: what should you do with this finding?]

### [Angle 2]
[Key findings, with sources]

**Verdict:** [Act now | Bookmark | Ignore | Needs more research]
[One sentence]

...

## Contradictions & Tensions

[Where did the research angles disagree?]

| Topic | Angle A Says | Angle B Says | Implication |
|-------|--------------|--------------|-------------|
| [topic] | [finding] | [conflicting finding] | [what this tension reveals] |

**Key insight from contradictions:**
[The disagreement between sources often reveals the most valuable insight.]

## Mission Relevance

[This section adapts to the member's mission context.]

- **[Member's niche]** — How does this apply to your specific coaching work?
- **[Their clients]** — What would [client avatar] need to know about this?
- **[Their delivery model]** — How could this change your [group/1:1/workshop] sessions?
- **[Their terminology]** — Does this connect to any of your named frameworks?

[If no mission context was loaded, use generic categories:]
- **Your coaching practice** — How does this apply to your work?
- **Your clients** — What would your clients need to know?
- **Your content** — What angles does this open for newsletters, posts, or videos?

## Action Verdicts Summary

| Finding | Verdict | Next Step |
|---------|---------|-----------|
| [Key finding 1] | Act now / Bookmark / Ignore | [Specific action] |
| [Key finding 2] | Act now / Bookmark / Ignore | [Specific action] |

## Learning Triggers

- **Questions raised:** [What this makes you wonder about]
- **Content angles:** [Story hooks or insights for your content]
- **Follow-up research:** [What would a second swarm investigate?]

## Sources
- [Source 1](URL)
- [Source 2](URL)

## Open Questions
[What we still don't know]

Phase 8: Save and Report

Step 1: Determine save location

1. If a vault/Zettelkasten was found in Phase 2 → save there as Research Swarm - [Topic] - YYYY-MM-DD.md
2. If no vault → save to project root ./research-reports/Research Swarm - [Topic] - YYYY-MM-DD.md (create directory if needed)
3. Offer: "Want me to save this somewhere else?"

Step 2: Link from daily note (optional)

1. Check if daily notes exist: Glob for today's date pattern (e.g., **/YYYY-MM-DD.md)
2. If found → offer to add to Captures section: - [[Research Swarm - Topic - YYYY-MM-DD]] - Brief summary
3. If not found → skip silently

Step 3: Report summary

Present the executive summary and action verdicts to the user.

When to Use This Skill

• Complex topics needing multiple perspectives
• Competitive research (what tools/approaches are available)
• Technology evaluation for your coaching practice
• Market research for content (newsletter, social, video)
• Any "research X" task that's not a quick lookup

Boundaries

Cost Optimization

• Sub-agents use haiku (fast, cheap for research gathering)
• Cap at 5 sub-agents per swarm
• The value is in the STRUCTURE of inquiry — multiple angles revealing what one perspective misses