| name | deep-research |
| description | Web research with Graph-of-Thoughts for fast-changing topics. Use when user requests research, analysis, investigation, or comparison requiring current information. Features hypothesis testing, source triangulation, claim verification, Red Team, self-critique, and gap analysis. Supports Quick/Standard/Deep/Exhaustive tiers. Creative Mode for cross-industry innovation. |
Deep Research
Enhanced research engine for topics where training data is outdated.
Quick Start
Standard Mode
CLASSIFY → LANDSCAPE SCAN → RECENCY PULSE → SCOPE → HYPOTHESIZE → PLAN → [PLAN PREVIEW*] → RETRIEVE
→ GAP ANALYSIS → TRIANGULATE → SYNTHESIZE → RED TEAM → SELF-CRITIQUE → PACKAGE
*Deep+ tier only
LANDSCAPE SCAN (MANDATORY - Before Anything Else)
[Search for OVERVIEW first - NO known entity names in query!]
WebSearch: "[topic] landscape overview [current year]"
WebSearch: "top [topic] list [current year]"
WebSearch: "[topic] ecosystem players [current year]"
❌ WRONG: "DeepSeek Qwen performance 2025" (uses names you already know)
✅ RIGHT: "China open source LLM models list 2025" (discovers what exists)
→ Extract ALL entity names from results
→ List: Discovered (new to you) vs Confirmed (you knew)
→ THEN proceed to RECENCY PULSE
Why: You cannot research what you don't know exists. Scan the landscape FIRST.
RECENCY PULSE (MANDATORY - After Landscape Scan)
[Search for LATEST news — within days/weeks, not just "this year"]
WebSearch: "[topic] latest news this week [current month] [current year]"
WebSearch: "[topic] new release announcement [current month] [current year]"
WebSearch: "[upstream provider 1] latest release [current year]"
WebSearch: "[upstream provider 2] latest release [current year]"
→ Check: anything released in the last 7-30 days?
→ If yes: add to entity list, flag as BREAKING/RECENT
→ THEN proceed to SCOPE with complete picture
UPSTREAM CHECK (part of Recency Pulse):
For any product/platform research, identify the SUPPLY CHAIN:
- Who MAKES the underlying technology? (e.g., OpenAI → GPT, Anthropic → Claude)
- Who DISTRIBUTES it? (e.g., Microsoft → Copilot, GitHub → Copilot)
- Who COMPETES with it? (e.g., Google → Gemini)
Search EACH upstream provider directly — don't rely on downstream announcements.
Example for "Microsoft Copilot":
Upstream: OpenAI (GPT models), Anthropic (Claude models)
Downstream: Microsoft (Copilot products)
→ Search "OpenAI latest model [month] [year]"
→ Search "Anthropic latest release [month] [year]"
→ Search "Microsoft Copilot new features [month] [year]"
Why: Downstream products lag behind upstream releases. A new model from
OpenAI/Anthropic may not appear in "Microsoft Copilot updates" for weeks.
If you only search downstream, you miss what's coming or just arrived.
Anti-pattern: ค้นแค่ "Microsoft Copilot new features 2026" แล้วหยุด
Better: ค้น upstream (OpenAI, Anthropic) + downstream (Microsoft) + "this week/month"
Creative Mode
ABSTRACT → MAP (3-5 domains) → SEARCH → GENERALIZE → SYNTHESIZE
Trigger: "creative mode", "cross-industry", "what do others do"
Example: "ทำยังไงให้คนมา engage กับ online course มากขึ้น?"
→ ABSTRACT: "retention + engagement ในกิจกรรมที่ทำซ้ำ"
→ MAP: Gaming (streaks, XP), Fitness apps (habit loops), YouTube (thumbnails, hooks), Loyalty programs (tiers)
→ SEARCH each domain → GENERALIZE patterns → SYNTHESIZE recommendations
Classification
| Type | When | Process | Example |
|---|
| A | Single fact | WebSearch → Answer | "Python 3.13 release date คือเมื่อไหร่?" |
| B | Multi-fact | Scan → Retrieve → Synthesize | "เปรียบเทียบ pricing ของ cloud GPU providers" |
| C | Judgment needed | Full 6 phases | "ควรใช้ Next.js หรือ Astro สำหรับ blog?" |
| D | Novel/conflicting | Full + Red Team | "AI จะแทนที่ data analyst ภายใน 3 ปีจริงไหม?" |
Intensity Tiers
| Tier | Sources | When |
|---|
| Quick | 5-10 | Simple question |
| Standard | 10-20 | Multi-faceted |
| Deep | 20-30 | Novel, high stakes |
| Exhaustive | 30+ | Critical decision |
Search & Evidence
Parallel Search (MANDATORY)
[Single message — always 2-3 queries at once]
WebSearch: "[topic] [current year]"
WebSearch: "[topic] limitations"
WebSearch: "[topic] vs alternatives"
Claim Types
| Type | Requirements | Example |
|---|
| C1 (Key claim) | Quote + 2+ sources + confidence | "Next.js มี market share 42%" |
| C2 (Supporting) | Citation required | "Vercel เป็นผู้พัฒนา Next.js" |
| C3 (Common knowledge) | Cite if contested | "React เป็น library ยอดนิยม" |
Confidence Format (C1 claims)
**Claim:** [Statement]
**Confidence:** HIGH/MEDIUM/LOW
**Reason:** [Why this confidence level]
**Sources:** [1][2]
Anti-Hallucination
- Every C1 cites [N] immediately
- Use "According to [1]..."
- Admit: "No sources found for X"
Research Sufficiency
"เมื่อไหร่ถึงจะพอ?"
| Signal | หมายความว่า |
|---|
| Saturation | 3 sources ต่อเนื่องไม่ให้ข้อมูลใหม่ → พอแล้ว |
| Convergence | หลาย sources สรุปเหมือนกัน → confidence สูง |
| Contradiction | Sources ขัดแย้งกัน → ต้อง dig deeper หรือ flag uncertainty |
| Diminishing returns | เพิ่ม search แต่ได้แค่ rephrase ของเดิม → หยุดได้ |
Quick tier: หยุดเมื่อ saturation
Standard: หยุดเมื่อ convergence + gap analysis ไม่เจอ gap สำคัญ
Deep/Exhaustive: หยุดเมื่อ Red Team challenge ไม่พบจุดอ่อนใหม่
Facilitation Guide
Progress Reporting
ทุกๆ 5-8 sources → update ผู้ใช้:
"สรุปที่พบจนถึงตอนนี้: [key findings]
ยังมีคำถามค้าง: [gaps]
จะ search ต่อเรื่อง [next direction] นะคะ"
When to Ask User
| สถานการณ์ | ถามว่า |
|---|
| Topic กว้างเกินไป | "อยากเน้นมุมไหนคะ? [option A] หรือ [option B]?" |
| เจอ sub-topic น่าสนใจ | "เจอเรื่อง X ที่เกี่ยวข้อง — อยากให้ขุดลึกไหมคะ?" |
| Sources ขัดแย้ง | "แหล่ง A บอกว่า X แต่แหล่ง B บอกว่า Y — พี่ระ lean ทางไหนคะ?" |
| Deep+ tier, plan ready | "นี่คือ plan สำหรับ research — approve ก่อนไปต่อนะคะ" |
Don't Ask — Just Do
- Type A questions → ตอบเลย
- Choosing search queries → ทำเลย ไม่ต้องถาม
- Formatting output → ใช้ template ได้เลย
Tools & Fallbacks
URL Fallback
If WebFetch returns 403:
curl -s --max-time 60 "https://r.jina.ai/https://example.com"
GitHub Repository Research
เจอ repo น่าสนใจ → ถาม user ก่อน clone:
"เจอ repo ที่น่าสนใจ: [repo-name] — ต้องการให้ clone มาศึกษา code ไหมคะ?"
If agreed:
mkdir -p /mnt/d/githubresearch && cd /mnt/d/githubresearch && git clone [repo-url]
Key files: package.json/pyproject.toml → src/ main logic → README.md
References
| Script | Purpose |
|---|
scripts/validate_report.py | 9-check quality validation |
Output File (MANDATORY)
After completing research, ALWAYS save to markdown file:
research/[topic-slug]-[YYYY-MM-DD].md
Example: research/china-opensource-ai-2025-01-04.md
- Create
research/ folder if it doesn't exist
- Why: Research takes effort. Save it for future reference.
Related Skills
/boost-intel — Apply critical thinking to research findings/generate-creative-ideas — Creative Mode for cross-industry innovation/skill-creator-thepexcel — Research domain expertise for skill creation/extract-expertise — Research to prepare expert interviews
