Deep Research

Produce a thorough, cited research report by decomposing a topic, searching multiple sources via firecrawl/exa MCP tools, deep-reading the strongest ones, and synthesizing findings with explicit source attribution.

Drift-prone surface. MCP tool names, quotas, and result shapes for firecrawl/exa change over time. Verify the configured MCP tools and current docs before promising coverage or quoting live source counts.

Tooling (MCP preferred, built-in fallback always available)

At least one of:

• firecrawl — firecrawl_search, firecrawl_scrape, firecrawl_crawl
• exa — web_search_exa, web_search_advanced_exa, crawling_exa

Both together give the best coverage; either alone is workable. Configure in ~/.claude.json or the equivalent MCP config.

Fallback tier — no MCP server configured. If neither firecrawl nor exa is available (check before starting, don't discover it mid-run), run the same workflow on the built-in WebSearch (breadth) and WebFetch (deep-reads) tools: same decomposition, same 15-30-source target, same deep-read-the-strongest-3-5 rule, same citation/confidence standards. The loss is crawl depth and structured extraction, not method — note "built-in tools, no crawl tier" in the report's methodology line so the reader can calibrate. Never dead-end on the missing servers.

Workflow

1. Clarify the goal. Ask 1-2 quick questions — "what's your goal: learning, deciding, or writing something?", "any specific angle or depth?" If the user says "just research it," proceed with reasonable defaults instead of blocking on this.

2. Decompose into 3-5 sub-questions. E.g. "impact of AI on healthcare" splits into: main applications today, measured clinical outcomes, regulatory challenges, leading companies, market size/trajectory.

3. Search each sub-question across sources. Use 2-3 keyword variations per sub-question, mixing general and news-focused queries. Target 15-30 unique sources total. Prioritize academic/official/reputable-news over blogs, and blogs over forums.

firecrawl_search(query: "<sub-question keywords>", limit: 8)
web_search_exa(query: "<sub-question keywords>", numResults: 8)
web_search_advanced_exa(query: "<keywords>", numResults: 5, startPublishedDate: "2025-01-01")

4. Deep-read the strongest 3-5 sources in full — don't synthesize from search snippets alone.

firecrawl_scrape(url: "<url>")
crawling_exa(url: "<url>", tokensNum: 5000)

5. Synthesize into a structured report:

# [Topic]: Research Report
*Generated: [date] | Sources: [N] | Confidence: [High/Medium/Low]*

## Executive Summary
[3-5 sentence overview of key findings]

## 1. [First Major Theme]
- Key point ([Source Name](url))
- Supporting data ([Source Name](url))

## 2. [Second Major Theme]
...

## Key Takeaways
- [Actionable insight 1]
- [Actionable insight 2]

## Sources
1. [Title](url) — [one-line summary]

## Methodology
Searched [N] queries across web and news. Analyzed [M] sources. Sub-questions: [list]

6. Deliver. Short topics: post the full report in chat. Long reports: post executive summary + key takeaways, save the full report to a file.

Parallelizing broad topics

For wide-scope topics, split sub-questions across parallel agents (e.g. via the Agent tool) — one agent per 1-2 sub-questions, each independently searching and reading, with the calling session synthesizing all returned findings into the final report.

Quality rules

1. Every claim needs a source — no unsourced assertions.
2. Cross-reference — a claim backed by only one source gets flagged as unverified.
3. Recency matters — prefer sources from the last 12 months unless historical context is the point.
4. Acknowledge gaps — if a sub-question turned up nothing solid, say so explicitly rather than papering over it.
5. No hallucination — "insufficient data found" beats a fabricated answer.
6. Separate fact from inference — label estimates, projections, and opinions as such.

Anti-pattern quick reference

Pre-flight checklist — run before delivering any research report

• Topic was decomposed into 3-5 concrete sub-questions before searching
• 15-30 unique sources were searched; 3-5 strongest were deep-read in full
• Every claim in the report traces to a cited source
• Single-source claims are flagged as unverified rather than stated as fact
• Gaps (sub-questions with weak/no evidence) are stated explicitly, not hidden
• Report includes Executive Summary, themed sections, Key Takeaways, Sources, and Methodology
• Delivery format (inline vs. file) matches report length