research-synthesizer（研究合成器）

Research Synthesizer Skill

Multi-source search → deduplicate → synthesize → cite. Concise answer under ~400 words, always.

When to Use

Trigger phrases:

• "research [topic]"
• "find out about [topic]"
• "what do you know about [topic]"
• "synthesize [topic]"

• "look up [topic]"

Step-by-Step Process

Step 0: Clarify the Brief

Before any research on companies, products, or competitors — ask or verify:

1. What is the positioning of OUR product? Don't assume. Ask if unclear.
2. What is the scope? Competitor analysis? Market sizing? Both?
3. What will the output be used for? Pitch deck? Internal doc? Strategy?

This prevents writing a wrong document that needs to be rewritten.

Step 0b: Question Decomposition (GPT Researcher Pattern)

Before searching, decompose the question into specific sub-questions:

Input: "What is Paperclip and how does it compare to monday.com?"

Sub-questions:
1. What is Paperclip? What does it do?
2. Who built it and when?
3. What are its core features?
4. How is it positioned vs. project management tools?
5. What does monday.com offer that Paperclip doesn't (and vice versa)?

Rule: For broad or multi-faceted questions (competitive analysis, "explain X", "compare A and B") — always decompose first. For simple factual questions ("who founded X", "when did Y happen") — skip this step.

Each sub-question becomes its own search query. This produces deeper, less biased results than 5 phrasings of the same question.

Step 1: Classify the Question

Before searching:

• Language: Is the question in Hebrew? → search in both Hebrew AND English
• Type: Factual? Opinion/trend? Technical? Recent event?
• Scope: Narrow (specific fact) or broad (overview topic)?

Adjust query phrasings accordingly.

Step 2: Generate Query Variants

Create 3–5 distinct query phrasings to maximize coverage and reduce bias:

Example — question: "What is LangGraph?"

• Q1: "What is LangGraph and how does it work"
• Q2: "LangGraph framework overview"
• Q3: "LangGraph tutorial explained"
• Q4: *(skip — English topic)*
• Q5: "LangGraph 2024 use cases"

Example — question: "What is LangGraph?"

• Q1: "What is LangGraph and how does it work"
• Q2: "LangGraph framework overview"
• Q3: "LangGraph explained simply"
• Q4: "LangGraph explained" (if topic has non-English coverage)
• Q5: "LangGraph 2025 latest"

Step 2b: Verify Companies — Visit Their Website First

MANDATORY for any competitor/company research:

Before writing anything about a company:

1. web_fetch their main URL (homepage + relevant sub-pages: /agents, /product, /pricing)
2. web_search "[company] funding 2026" AND "[company] review 2026"
3. Only write what you actually found. If unverified — say "unverified"

Never assume a company's capabilities from its category name. Example: "issue tracker" does NOT mean "no agents." Verify.

Step 3: Run Searches (Parallel)

Run all query variants using web_search. Collect:

• Title
• URL
• Snippet

Do not fetch full page content unless snippet is insufficient.

Step 4: Deduplicate, Filter & Score Sources

From all results:

1. Remove duplicate URLs
2. Remove results that don't address the question
3. Remove results older than 2 years for fast-moving topics (AI, tech, news)
4. Score source credibility:

- High: Official docs, peer-reviewed, major publications (TechCrunch, Wired, HBR) - Medium: Reputable blogs, GitHub repos, well-known newsletters - Low: Forums, anonymous posts, marketing pages

1. Prioritize high-credibility sources. If only low-credibility sources available — flag it.

Target: 5–10 sources for deep research, 3–5 for quick questions.

Step 5: Synthesize

Write the answer in this format:

[3–5 sentence summary that directly answers the question]

Key points:
• [point 1]
• [point 2]
• [point 3]
• [point 4 — optional]

Sources:
1. [Title] — [URL]
2. [Title] — [URL]
3. [Title] — [URL]

Synthesis rules:

• Lead with the direct answer, not background
• Use the user's language (Hebrew question → Hebrew answer)
• Max ~400 words total
• If sources conflict → note the disagreement briefly
• If information is outdated → flag it

Step 6: Deliver

Send the synthesized answer. Do NOT:

• Dump raw search results
• List every source found
• Add filler like "Great question!" or "I searched the web for..."
• Pad with unnecessary caveats

Output Format Template

🔍 [Topic]

[Direct 3-5 sentence answer]

📌 Key Points:
• ...
• ...
• ...

📚 Sources:
1. [Title] — [URL]
2. [Title] — [URL]
3. [Title] — [URL]

Example

Input: "Research: What is Model Context Protocol?"

Output:

🔍 Model Context Protocol (MCP)

Model Context Protocol (MCP) is an open standard developed by Anthropic that lets LLMs connect uniformly to tools, APIs, and external data sources. Instead of each integration requiring custom code, MCP defines a shared language between the model and the tool server.

📌 Key Points:
• Client-server protocol: the LLM is the client, tools are servers
• Supports stdio and HTTP transport
• Enables: tool calling, resource access, prompts
• Widely adopted: Claude, Cursor, VS Code, and more
• Open source — SDK available for Python, TypeScript, Java

📚 Sources:
1. MCP Official Docs — https://modelcontextprotocol.io
2. Anthropic MCP Announcement — https://www.anthropic.com/news/model-context-protocol
3. MCP GitHub — https://github.com/modelcontextprotocol

Hebrew Search Strategy

For Hebrew questions, always search in both languages:

If the topic is inherently Israeli (local news, Israeli law, etc.) → weight Hebrew sources more.

Rules

1. Always cite sources — no answer without at least 2 URLs. For competitive analysis: minimum 5 sources.
2. Clarify positioning before writing (Step 0) — especially for competitive analysis. Ask what OUR product does before comparing.
3. Verify companies from their own website (Step 2b) — never assume from category name.
4. Deep questions → decompose first (Step 0b). Simple facts → skip decomposition.
5. Max ~400 words — be concise, not exhaustive
6. One clean doc, not multiple drafts — get it right before publishing
7. Direct answer first — no preamble, no "I will now search..."
8. Hebrew in, Hebrew out — match the user's language
9. Flag uncertainty — if sources conflict or data is stale, say so
10. No raw dumps — synthesize, don't copy-paste snippets
11. React 👍 when owner requests research, ✅ when delivered
12. After delivering research — write summary to memory/whatsapp/dms/<PHONE-sanitized>/context.md if topic was important

Cost Notes

• 3–5 web_search calls per research request — moderate cost
• Avoid web_fetch unless snippets are truly insufficient
• For simple factual questions (capital cities, dates, etc.) → single search is enough, skip full synthesizer flow
• Cache: if the same topic was researched in the last hour, reuse results