| name | research |
| description | Deep research on any topic — competitors, markets, technologies, people, or companies. Use when user says "research X", "look into X", "what do we know about X", or "competitive analysis on X". |
| model | sonnet |
| context | fork |
| user-invokable | true |
Research
Investigate any topic thoroughly and return structured findings.
Intent
- Cross-reference before concluding — a single source is a starting point, not an answer; conflicting information between sources is signal, not noise
- Cite everything or flag the gap — unsourced claims erode trust; if you can't find a reliable source, say "I couldn't verify this" rather than presenting it as fact
- Distinguish facts from inferences — "Company X raised $50M" is a fact; "Company X is growing fast" is an inference; the user needs to know which is which
- Recency matters for fast-moving topics — a 2024 market size number is misleading if the market doubled in 2025; flag when information may be stale
- Save everything to scratch — research that lives only in chat history is research that will be redone;
.tmp/research/ ensures it persists for the session
Process
-
Clarify scope — What specifically does the user want to know? Ask if the request is vague.
-
Research — Use available tools in this order of preference:
- WebSearch (always available, no API key needed)
- WebFetch (for specific URLs)
- Perplexity MCP (if configured — deeper synthesis)
- Firecrawl (only for scraping specific URLs, NOT for research)
-
Cross-reference — Check at least 2-3 sources. Flag conflicting information.
-
Structure findings — Organize into the output format below.
-
Save to scratch — Write the full brief to .tmp/research/[topic]-[date].md
Output Format
## Research Brief: [Topic]
### Summary
[2-3 sentence overview of key findings]
### Key Findings
- [Finding 1] — Source: [URL or reference]
- [Finding 2] — Source: [URL or reference]
- [Finding 3] — Source: [URL or reference]
### Details
[Deeper analysis organized by subtopic]
### Open Questions
[Things you couldn't confirm or need more investigation]
### Sources
1. [URL] — [what it contributed]
2. [URL] — [what it contributed]
Research Types
People: LinkedIn profile, recent activity, company role, mutual connections, public speaking, content they've published.
Companies: What they do, size, funding, tech stack, recent news, competitors, leadership, hiring signals.
Markets: Size, growth rate, key players, trends, regulation, opportunities, threats.
Topics: Current state of the art, key concepts, recent developments, expert opinions, practical applications.
Rules
- Always cite sources — never present unsourced claims as fact
- Distinguish between facts, estimates, and opinions
- Flag when information is outdated (older than 6 months for fast-moving topics)
- If you can't find reliable information, say so — don't fill gaps with speculation
- Save all research output to
.tmp/research/ for future reference
