Firecrawl Deep Research

Use this only for report-scale research: a rigorous, cited synthesis the user

explicitly wants delivered as a formal written report. If the request is a

product pick, a top-N list, a quick lookup, or anything answerable with a short

search, stop; do not use this skill, let the request be handled the standard

way.

This skill gathers its evidence from the open web. If the evidence base is the

published literature — a literature review, or a biomedical, clinical, or other

scientific topic where the answer lives in papers — use

firecrawl-research-papers instead; it

queries Firecrawl's paper index rather than searching websites.

Onboarding Interview

Infer the topic and output format from context. Before starting, unless already specified, always ask one short question to define the scope:

"How long do you want this research task to run?"

Map the answer to a depth tier in the Collection Plan below:

• A few minutes → Quick

• ~10-15 minutes → Thorough

• Longer / no limit → Exhaustive

If the topic itself is unclear, you may ask at most 1-2 additional concise

questions (topic, or a critical angle/source constraint). Otherwise proceed once

the runtime is set.

Firecrawl Collection Plan

Use Firecrawl search and scrape through the CLI or equivalent tool surface. Match

depth to the runtime the user chose during onboarding.

• Quick (~a few minutes): search 3-5 queries and scrape 5-10 high-quality sources.

• Thorough (~10-15 minutes): search 5-10 queries from different angles and scrape 15-25 sources.

• Exhaustive (longer): search 10+ queries and scrape 25+ sources, including primary sources, research papers, expert views, and contrarian sources.

Avoid re-scraping URLs already returned with full content from a search-with-scrape result.

When Published Papers Are The Evidence

Search and scrape reach web pages. They do not query Firecrawl's research paper

index, which holds paper abstracts with full text reachable per paper — largely

biomedical and life-science literature from PubMed, bioRxiv, and medRxiv, plus

arXiv preprints in CS, physics, and math.

Hand off to firecrawl-research-papers

when the report's evidence base is the published literature — a biomedical,

clinical, drug, gene, disease, epidemiology, or public-health topic, or any

request phrased as a literature review, systematic review, or survey of studies.

That skill uses firecrawl_research_* (MCP) / firecrawl research (CLI) to

search abstracts, expand to related papers, and verify claims inside a paper

body — none of which plain search and scrape can do.

If the report needs both — the literature and market, policy, or news context —

run the paper work through that skill and keep the web collection above for the

rest, then synthesize here.

Note that passing categories: ["research"] to Firecrawl search does not query

the paper index either. It filters an ordinary web search to research-affiliated

websites — the list includes PubMed, bioRxiv, medRxiv, arXiv, and publisher

sites — and returns their web pages, not the paper records behind them.

Parallel Work

If appropriate, use sub-agents or equivalent parallel task runners by research angle:

• overview and definitions

• technical or implementation details

• market and industry context

• contrarian views, risks, and limitations

• primary sources and official docs

Each researcher should return claims, source URLs, source quality notes, and uncertainty.

Final Deliverable

Default structure:

# Deep Research: [Topic]

## Executive Summary

[2-3 paragraphs]

## Key Findings

[Numbered findings with source links]

## Detailed Analysis

[Themes, evidence, and synthesis]

## Contrarian Views And Risks

[Counterarguments, limitations, failure modes]

## Open Questions

[What remains uncertain]

## Sources

[Every URL used with a one-line note]

## Rerun Inputs

workflow: firecrawl-deep-research

topic: [topic]

depth: [quick/thorough/exhaustive]

output: [markdown/json/brief]

Quality Bar

• Cite sources for factual claims.

• Prefer primary sources when available.

• Flag uncertainty and conflicting evidence.

• Synthesize instead of listing scrape summaries.