research-orchestrator

Research Orchestrator

Systematic multi-workstream research for any question. Decompose → Execute → Synthesize.

---

Phase 1: Plan

Before searching anything:

1. Understand the question. What exactly needs to be learned? What decision or output will this research inform?

2. Decompose into 2–4 workstreams. Each workstream must have:

   • A clear, non-overlapping scope
   • 3–5 specific sub-questions or sections to answer
   • An implied search strategy (what sources, what angles)

3. Present the plan to the user in this format:

## Research Plan: [Topic]

**Core question:** [What we're actually trying to answer]

### Workstream 1: [Name]
**Scope:** [1 sentence]
**Key questions:**
1. [Question]
2. [Question]
3. [Question]

### Workstream 2: [Name]
...

### Synthesis
After all workstreams: cross-cutting themes, contradictions, confidence assessment, recommendations.

---
Proceed with this plan? Or adjust workstream scope?

4. Wait for user confirmation before executing. Do not begin research until approved.

---

Phase 2: Execute Workstreams

Execute workstreams sequentially (Claude.ai constraint — no parallel agents). For each workstream:

Search Protocol

For each sub-question in the workstream:

1. Run 1–3 targeted web searches (keep queries 2–5 words, specific)
2. Fetch full content from the most relevant result(s) using web_fetch
3. Extract and record findings before moving to the next search
4. Never run 3+ searches in a row without recording findings — research loops produce nothing

Source Standards

• Every quantitative claim requires an inline source
• Prefer primary sources (company filings, peer-reviewed papers, official data) over aggregators
• If sources conflict, note the conflict — don't pick one arbitrarily
• Flag any finding where source quality is low or data is >2 years old

Depth Calibration

• Simple factual question: 2–3 searches per workstream, ~200–400 words output
• Strategic/analytical question: 4–8 searches per workstream, ~500–1,000 words output
• Deep research request: 8–15 searches per workstream, 1,000–2,000 words output

Progress Reporting

After completing each workstream, briefly report:

✓ Workstream 1 complete: [2-sentence summary of what was found]
Starting Workstream 2...

---

Phase 3: Synthesize

After all workstreams complete, produce a synthesis document with these sections:

1. Executive Summary

3–5 bullets. The most important things learned. What would change decisions.

2. Key Findings by Theme

Organized by cross-cutting theme — NOT by workstream. This is the value-add: insights that span multiple workstreams, patterns that only emerge when the full picture is assembled.

3. Contradictions & Tensions

What did different workstreams find that conflicts? Where is the evidence genuinely uncertain? Do not paper over these — flag them explicitly.

4. Confidence Assessment

Finding	Confidence	Basis
[Claim]	High / Medium / Low	[Why]

High = multiple corroborating primary sources Medium = single good source or multiple secondary sources Low = limited data, old data, or conflicting sources

5. Recommended Next Steps

What should the user do with this research? What gaps remain? What would move low-confidence findings to high?

---

Key Rules

1. Never start research without user approval of the plan. The decomposition IS the insight — wrong workstreams waste everything that follows.

2. Write-then-search, not search-then-search. After every 1–2 searches, record findings. Never queue up 5 searches and then try to synthesize from memory.

3. Contradictions are features, not bugs. If workstreams surface conflicting data, report it. The user hired a researcher, not a PR firm.

4. Depth matches the question. Don't produce a 3,000-word synthesis for a question that has a clear 200-word answer. Calibrate.

5. Workstream scope should be exhaustive but non-overlapping. If two workstreams would both naturally search for the same thing, adjust the split before launching.

6. Flag source quality inline. If you're citing a Reddit post or an SEO farm, say so. The user needs to know where the evidence is solid.

7. The synthesis is not a summary. It surfaces cross-cutting patterns, not a workstream-by-workstream recap. If the synthesis just repeats what each workstream said, it failed.

---

Workstream Decomposition Patterns

Use these as starting points — adapt to the actual question:

Competitive / Market Analysis

• W1: Market structure & sizing
• W2: Competitor positioning & differentiation
• W3: Customer behavior & unmet needs
• W4: Regulatory / macro environment

Investment / Deal Due Diligence

• W1: Asset / business fundamentals
• W2: Market comparable & valuation
• W3: Risk factors & downside scenarios
• W4: Operator / management track record

Technology Evaluation

• W1: Technical capabilities & architecture
• W2: Adoption landscape & case studies
• W3: Limitations, failure modes & alternatives
• W4: Vendor / ecosystem health

Person / Organization Research

• W1: Background & track record
• W2: Public reputation & third-party assessments
• W3: Network, affiliations & interests
• W4: Recent activity & stated positions

Strategic Decision Support

• W1: Factual landscape (what is true now)
• W2: Expert opinion & practitioner experience
• W3: Historical precedents & analogues
• W4: Dissenting views & counterarguments

---

Output Format Options

Ask the user (or infer from context) which output format they need:

• Conversational: Findings presented as prose in the chat. Good for quick research, single-question answers.
• Structured report: Synthesis document formatted with headers, tables, confidence ratings. Good for sharing, decision support, long shelf life.
• Actionable brief: Executive summary + top 3 recommendations only. Good for busy decisions.

Default to structured report unless the question is simple or the user signals otherwise.