research-orchestrator

Files (3)

• README.md
• SKILL.md
• hero.png · assets/

SKILL.md (first lines)

# Research Orchestrator

Multi-perspective research using Fan-Out / Fan-In pattern. Spawns 5 parallel sub-agents to explore different axes of a research query, then synthesizes findings with a high-reasoning agent.

## How It Works

```
Query → 5 Sonnet agents (parallel) → accumulate in ~/active/ → deduplicate → 1 Opus synthesizer → final report
```

**Options:**

- `--quick` — Early stopping: return results as soon as 3 of 5 agents complete (faster, less comprehensive)
- (default) — Wait for all 5 agents before synthesis

## Step 1: Receive and Decompose the Query

When the user provides a research query (either as an argument or via prompt):

1. If no query was provided as an argument, ask for one using AskUserQuestion.
2. Decompose the query into **5 research axes**. Choose axes that maximize coverage with minimal overlap.

**Axis selection heuristic** — pick 5 from these common lenses (adapt to the query):

| Axis Type                    | What It Covers                                           |
| ---------------------------- | -------------------------------------------------------- |
| **Technical/How**            | Implementation details, architecture, tools, specs       |
| **Market/Who**               | Players, competitors, market share, adoption             |
| **Trends/When**              | Timeline, momentum, recent developments, trajectory      |
| **Risks/Why Not**            | Limitations, failures, criticisms, trade-offs            |
| **Opportunities/Why**        | Use cases, benefits, untapped potential, synergies       |
| **Community/Social**         | What practitioners say, sentiment, real-world experience |
| **Historical/Context**       | How we got here, evolution, prior art                    |
| **Regulatory/Legal**         | Compliance, policy, governance, ethical considerations   |
| **Economic/Cost**            | Pricing, ROI, total cost, funding landscape              |
| **Comparative/Alternatives** | Alternative approaches, competing paradigms              |

3. Present the 5 chosen axes to the user before launching agents. Example:

```
Researching: "Best approaches to building AI agents in 2026"

Axes:
1. Technical architectures & frameworks
2. Market landscape & key players
3. Real-world practitioner experiences & pitfalls
4. Cost structures & scaling economics
5. Emerging trends & trajectory (last 6 months)

Launching 5 research agents...
```

## Step 2: Fan-Out — Launch 5 Parallel Research Agents

Create a timestamped accumulation file:

```bash
# Format: ~/active/research-YYYY-MM-DD-HHMMSS.md
```

Write the file header:

```markdown
# Research: {query}

**Date:** {date}
**Axes:** {list of 5 axes}

---
```

Then spawn **5 agents in a single message** (critical for parallelism) using the Agent tool:

- **model:** `sonnet` for each sub-agent
- **subagent_type:** `general-purpose`
- **run_in_background:** `true` for all 5

Each agent's prompt MUST include:

- The full research query for context
- Its specific axis assignment
- Instruction to use WebSearch and WebFetch for current information
- Instruction to write findings to the accumulation file using a clearly labeled section
- Instruction to include source URLs where possible
- Instruction to note confidence level (high/medium/low) for each finding
- The exact file path to write to

**Agent prompt template:**

```
You are a research agent investigating one axis of a larger research question.

RESEARCH QUERY: {query}
YOUR AXIS: {axis_name} — {axis_description}

Instructions:
1. Use WebSearch and WebFetch to find current, authoritative information about this axis.
2. Search for 3-5 different sub-queries to get broad coverage within your axis.
3. Write your findings to {file_path} by APPENDING (not overwriting) a section.
4. Format your section as:

## Axis {N}: {axis_name}

### Key Findings
- Finding 1 [confidence: high/medium/low]
  - Source: {url}
- Finding 2 ...

### Analysis
{2-3 paragraph analysis of what you found}

### Surprises or Outliers
{Anything unexpected or contrarian}

---

5. Be thorough but concise. Focus on facts, data, and expert opinions over generic summaries.
6. If you find conflicting information, note the conflict explicitly — include both sides and the source for each.
7. Tag each finding with a unique ID (e.g., A1-F1, A1-F2) so the synthesizer can track cross-axis overlaps.

From kabirdos/claude-toolkit, MIT.

More from kabirdos/claude-toolkit