| name | deep-research |
| description | Run deep research on a topic. Web searches, data collection, source checking, and a structured report. Use when the user needs more than a quick answer. |
| allowed_tools | ["WebSearch","WebFetch","Read"] |
Deep Research
Go deep on one topic. Come back with facts, not fluff.
When to Use
- User needs real data on a topic
- Writing a research report or blog post
- Need to understand a new space
- Preparing materials for investors or partners
Process
1. Scope
Ask or confirm:
- What's the question?
- How deep? (quick scan vs full report)
- Any angle? (tech, business, user behavior, etc.)
- Who's the audience? (self, team, investors, public)
2. Research
Run multiple web searches. Look for:
- Industry reports and data
- Academic papers or studies
- News articles (last 12 months)
- Expert opinions and blog posts
- Reddit/HN/Twitter discussions
- Company blogs and case studies
3. Verify
- Cross-check numbers across sources
- Flag conflicting data
- Note the source quality (press release vs research paper)
- Check dates. Flag anything older than 2 years.
4. Synthesize
Don't just list what you found. Connect the dots:
- What do the facts add up to?
- What's the pattern?
- What's missing from the data?
- What surprised you?
5. Write Report
Structure:
- TL;DR - 3 sentences max
- Background - Context someone needs to understand
- Key Findings - The meat. Numbered, sourced.
- Analysis - What it means. Your take.
- Open Questions - What we still don't know
- Sources - Every link, with date and source name
Rules
- Every claim needs a source
- Say "I couldn't find data on X" instead of making stuff up
- Include views that disagree with each other
- Keep your opinion in the Analysis section, not the Findings
- If the user asks for quick research, skip steps 3-4
Gotchas
- Old data kills credibility. A 2021 market size report is useless in 2026. Always check the publish date. Flag anything older than 18 months.
- Press releases aren't research. Company announcements are marketing. Cross-check with third-party sources, financial filings, or user data.
- Conflicting sources are a feature, not a bug. When two reports disagree, that's where the interesting analysis lives. Don't just pick one.
- "I couldn't find data" is a valid finding. Don't fill gaps with guesses. If the data doesn't exist, say so -- that's useful info for the user.
- Synthesis > summary. Listing 10 findings isn't research. Connect the dots -- what patterns emerge? What's the "so what?"
Output
Save to the project's 02-research/ folder.
