| name | deep-research |
| description | Deep multi-round web research with query decomposition, iterative refinement, source triangulation, and structured synthesis. Use when the user asks to research a topic in depth, investigate a complex question, compare multiple options, produce a research report, or needs comprehensive information from multiple sources. |
| license | MIT |
| allowed-tools | web_search web_fetch user_message |
Deep Research Methodology
You are conducting deep research — not a quick lookup. Your goal is to produce a thorough, well-sourced analysis that the user can trust and act on.
Phase 1: Decompose the Question
Before searching, break the user's question into 3–6 sub-questions that together cover the full scope. Write them out explicitly.
Example — user asks "Should we migrate from REST to GraphQL?":
- What are the current technical limitations of REST for our use case?
- What performance and developer-experience benefits does GraphQL offer?
- What are the known operational costs and pitfalls of GraphQL at scale?
- What do teams who migrated back from GraphQL report?
- What is the current industry adoption trend and tooling maturity?
This decomposition guides all subsequent searches.
Phase 2: Broad Search (Round 1)
For each sub-question, run a targeted web_search query. Use diverse query formulations:
- Factual query — direct question phrased for search engines
- Authoritative query — target official docs, research papers,
.gov, .edu, .org
- Contrarian query — "problems with X", "X criticism", "X vs Y disadvantages"
- Recent query — append current year or "2025" / "2026" for fast-moving topics
Request 5 results per query. Track all URLs seen — discard duplicates across queries.
After completing all broad searches, use user_message to send the user a brief progress update (e.g., "Completed broad search across 5 sub-questions. Found 18 unique results. Moving to deep dive on top 5 sources.").
Phase 3: Deep Dive (Round 2)
From Round 1 results, select the top 3–5 most promising URLs and fetch their full content with web_fetch. Prioritize:
- Primary sources (official documentation, research papers, data sets)
- In-depth technical posts with benchmarks or case studies
- Sources that represent opposing viewpoints
When reading fetched content:
- Extract specific data points: numbers, dates, benchmarks, quotes
- Note the publication date and author credentials
- If a page fails to load or is paywalled, note it and search for an alternative
Phase 4: Gap Analysis & Iterative Refinement (Round 3+)
After Round 2, review what you have and identify gaps:
- Are any sub-questions still poorly answered?
- Did new information reveal sub-questions you hadn't considered?
- Are there conflicting claims that need a tiebreaker source?
Run additional targeted searches to fill gaps. Repeat until each sub-question has at least 2 independent, credible sources — or you've exhausted reasonable search avenues.
Use user_message to update the user on gap analysis progress (e.g., "Gap analysis: 2 sub-questions need more evidence. Running targeted follow-up searches.").
Stop condition: You have converging evidence for each sub-question, OR you have clearly identified where evidence is lacking.
Phase 5: Source Evaluation
Before including any source in your report, assess it:
| Criterion | Strong | Weak |
|---|
| Type | Primary source, peer-reviewed, official docs | Aggregator, opinion blog, anonymous forum |
| Recency | Published within 1–2 years (for tech/policy topics) | Older than 3 years on a fast-moving topic |
| Independence | No financial stake in the claim | Vendor marketing, sponsored content |
| Consensus | 3+ independent sources agree | Single source, no corroboration |
| Specificity | Concrete data, benchmarks, examples | Vague claims, no supporting evidence |
Discard sources that score "Weak" on 3+ criteria unless they provide a unique perspective worth noting.
Phase 6: Synthesize & Report
Structure your final output as:
## Summary
[3–5 sentence executive summary answering the user's core question]
## Key Findings
### [Sub-question 1]
- [Finding with specific data and source attribution]
- [Finding with specific data and source attribution]
### [Sub-question 2]
- ...
(repeat for each sub-question)
## Conflicting Information
- [Source A claims X (reason/evidence), while Source B claims Y (reason/evidence)]
- [Your assessment of which is more credible and why]
## Gaps & Limitations
- [What couldn't be reliably determined]
- [Areas where evidence is thin or outdated]
## Confidence Assessment
[High / Medium / Low] — with a 1-sentence justification
## Sources
1. [Title] — [URL] — [Date] — [Used for: which finding]
2. ...
Rules
- Minimum 3 search rounds before producing a report (broad → deep → gap-fill)
- Never cite a source you haven't fetched and read — search snippets are unreliable
- Never omit contradictory evidence — report disagreements honestly
- Never present a single source as definitive — triangulate
- Always distinguish facts from opinions — "According to [source]..." vs stating as fact
- Flag uncertainty explicitly — "Evidence is limited" is more useful than false confidence
- Cite every claim — no unsourced assertions in the findings
