Research

Research

Produce thorough, cited research reports from multiple web sources — using exa (search) and firecrawl (scraping) MCP tools when available, otherwise the fallbacks below.

When to Activate

• User asks to research any topic in depth
• Competitive analysis, technology evaluation, or market sizing
• Due diligence on companies, investors, or technologies
• Any question requiring synthesis from multiple sources
• User says "research", "deep dive", "investigate", or "what's the current state of"

Backends — check availability before planning, not mid-run

Preferred, at least one of:

• exa — the search backend: web_search_exa, web_fetch_exa (the configured server exposes exactly these two; verified signatures live in the search-routing skill). It scored best in a 2026-09-13 blind eval, including on Chinese and site: queries.
• firecrawl — the scraper: firecrawl_scrape for JS-heavy pages or structured extraction, firecrawl_crawl for a site section. firecrawl_search is a fallback only — Free plan, 1,000 credits/month; on 402 / Insufficient credits stop calling it for the rest of the run and tell every subagent so.

Both together give the best coverage. Configure them as MCP servers (claude mcp add, or the mcpServers block of your Claude Code config / ~/.codex/config.toml for Codex). MCP servers are scoped per config directory — if you run more than one, a server added under one is absent under the other, so verify from the session you actually intend to research in.

Check first: are the tools present in this session? If neither backend is, do NOT plan a workflow around tools that will not answer. Fall back in this order:

1. /deep-research — the bundled dynamic workflow. It already does fan-out → fetch → cross-check → cited report on top of the built-in WebSearch, which is closer to this skill's output contract than anything you would improvise. Prefer it whenever the ask is a full multi-source report.
2. Built-in WebSearch / WebFetch — always available, no configuration. Use for narrower questions, and run the workflow below by hand: same sub-questions, same deep-read step, same citation rules, substituting WebSearch for *_search and WebFetch for *_scrape / web_fetch_exa.
3. Any search CLI or skill your environment provides — if one is installed and covers more sources than WebSearch, use it and say which.

Whichever you land on, name it before searching and record it in the report's Methodology section — coverage is not interpretable without it.

Workflow

Step 1: Understand the Goal

Ask 1-2 quick clarifying questions:

• "What's your goal — learning, making a decision, or writing something?"
• "Any specific angle or depth you want?"

If the user says "just research it" — skip ahead with reasonable defaults.

Step 2: Plan the Research

Break the topic into 3-5 research sub-questions. Example:

• Topic: "Impact of AI on healthcare"
  • What are the main AI applications in healthcare today?
  • What clinical outcomes have been measured?
  • What are the regulatory challenges?
  • What companies are leading this space?
  • What's the market size and growth trajectory?

Scenario checklists — when the topic matches one of these modes, make sure the sub-questions collect:

• Investor / fund diligence: fund size, stage, and typical check size; relevant portfolio companies; public thesis and recent activity; concrete fit / no-fit reasons; obvious red flags or mismatches.
• Competitive analysis: product reality, not marketing copy; funding and investor history if public; traction metrics if public; distribution and pricing clues; strengths, weaknesses, and positioning gaps.
• Market sizing: top-down estimates from reports or public datasets; bottom-up sanity checks from realistic customer-acquisition assumptions; an explicit assumption stated for every leap in logic.
• Technology / vendor evaluation: how it works; trade-offs and adoption signals; integration complexity; lock-in, security, compliance, and operational risk.

Step 3: Execute Multi-Source Search

For EACH sub-question, search using available MCP tools:

With exa (default):

web_search_exa(query: "<sub-question as a description of the ideal page>", numResults: 8,
               objective: "<which sources should rank first / be excluded>")

Add site:<domain> to query to stay on one site (verified to work on the remote exa server, 2026-09-15).

With firecrawl (fallback, only if exa comes back thin and credits remain):

firecrawl_search(query: "<sub-question keywords>", limit: 8)

Exa has no date-filter parameter. For recency-sensitive sub-questions, put the time constraint in query or objective (e.g. " 2025", "published after June 2025"); only reach for firecrawl_search with a date operator (e.g. <keywords> after:2025-01-01) when exa's results are clearly stale.

Search strategy:

• Use 2-3 different keyword variations per sub-question
• Mix general and news-focused queries
• Aim for 15-30 unique sources total
• Prioritize: academic, official, reputable news > blogs > forums

Step 4: Deep-Read Key Sources

For the most promising URLs, fetch full content:

With exa (default):

web_fetch_exa(urls: ["<url>"], maxCharacters: 20000)

With firecrawl:

firecrawl_scrape(url: "<url>", formats: ["markdown"])

web_fetch_exa handles static pages; if it fails, try the free curl -s https://r.jina.ai/<url> before spending credits. For JS-heavy pages, prefer firecrawl_scrape (keep formats: ["markdown"] — JSON formats cost +4 credits per page); to sweep a whole site section, use firecrawl_crawl.

Read 3-5 key sources in full for depth. Do not rely only on search snippets.

Step 5: Synthesize and Write Report

Structure the 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]
[Findings with inline citations]
- Key point ([Source Name](url))
- Supporting data ([Source Name](url))

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

## 3. [Third Major Theme]
...

## Key Takeaways
- [Actionable insight 1]
- [Actionable insight 2]
- [Actionable insight 3]

## Sources
1. [Title](url) — [one-line summary]
2. ...

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

Step 6: Deliver

• Short topics: Post the full report in chat
• Long reports: Post the executive summary + key takeaways, save full report to a file

Parallel Research with Subagents

For broad topics, use the Agent tool to parallelize:

Launch 3 research agents in parallel:
1. Agent 1: Research sub-questions 1-2
2. Agent 2: Research sub-questions 3-4
3. Agent 3: Research sub-question 5 + cross-cutting themes

Each agent searches, reads sources, and returns findings. The main session synthesizes into the final report.

Quality Rules

1. Every claim needs a source. No unsourced assertions.
2. Cross-reference. If only one source says it, flag it as unverified.
3. Recency matters. Prefer sources from the last 12 months.
4. Acknowledge gaps. If you couldn't find good info on a sub-question, say so.
5. No hallucination. If you don't know, say "insufficient data found."
6. Separate fact from inference. Label estimates, projections, and opinions clearly.

Examples

"Research the current state of nuclear fusion energy"
"Deep dive into Rust vs Go for backend services in 2026"
"Research the best strategies for bootstrapping a SaaS business"
"What's happening with the US housing market right now?"
"Investigate the competitive landscape for AI code editors"