Standard Mode: Enhanced Research Pipeline

Standard Mode: Enhanced Research Pipeline

Contents

• Phase 0: CLASSIFY
• Phase 0.5: DISCOVERY
• Phase 1: SCOPE
• Phase 1.5: HYPOTHESIZE
• Phase 2: PLAN
• Phase 2.5: PLAN PREVIEW
• Phase 3: RETRIEVE
• Phase 3.5: GAP ANALYSIS
• Phase 4: TRIANGULATE
• Phase 5: SYNTHESIZE
• Phase 6: RED TEAM
• Phase 6.5: SELF-CRITIQUE
• Phase 7: PACKAGE
• Quality Checklist

Phase 0: CLASSIFY

Objective: Route questions to appropriate process intensity

Gate: Classification explicit before proceeding.

Phase 0.5: LANDSCAPE SCAN

Objective: Map the full landscape BEFORE diving into specifics. Discover what exists.

Rationale: You cannot research what you don't know exists. Claude's training data has a cutoff. For fast-changing fields (AI, crypto, tech), new entities exist that Claude doesn't know about. Scanning the landscape FIRST prevents blind spots.

CRITICAL RULE: No Known Names in Scan Queries

❌ WRONG: "DeepSeek Qwen performance comparison 2025"
   → You're only searching things you already know!

✅ RIGHT: "China open source LLM models list 2025"
   → You're discovering what's out there

When to Apply

ALL research types — MANDATORY before SCOPE:

Process

1. Landscape Search (parallel, NO known entity names, use current year):

   WebSearch: "[topic] landscape overview [current year]"
   WebSearch: "top [topic] list [current year]"
   WebSearch: "[topic] ecosystem all players [current year]"
   WebSearch: "complete list [topic] [current year]"
   

2. Extract ALL Entities from Results:

   **Discovered (NOT in my training data):**
   - [Entity 1]: Brief description from search
   - [Entity 2]: Brief description from search
   
   **Confirmed (matches my knowledge but updated):**
   - [Entity A]: What's new
   - [Entity B]: Still relevant
   

3. Build Complete Picture:

   • Count total entities found
   • Note which are new vs known
   • Use discovered names in SCOPE and PLAN queries
   • Flag unknown entities for deeper investigation

Example

Topic: "Chinese open-source AI models"

Landscape Scan Queries (CORRECT):

"China open source AI models complete list 2025"
"Chinese AI ecosystem all players 2025"
"top Chinese LLM image video models 2025"

NOT This (WRONG - uses known names):

"DeepSeek vs Qwen comparison 2025" ← Skip this until AFTER scan!

After Landscape Scan:

Discovered: Wan 2.1, Kolors 2.1, MiniCPM4, Qwen3,
           DeepSeek-V3.2, HunyuanVideo-I2V, Step-2, Emu3
Confirmed: DeepSeek, Qwen, Yi, Baichuan, GLM

→ NOW you have complete picture for SCOPE

Anti-Patterns

Gate: Complete entity list extracted; landscape mapped before SCOPE.

Phase 1: SCOPE

Objective: Define research boundaries and success criteria

THINK Step: Take Stock (T)

Before searching, clarify prior knowledge. See researcher-thinking.md.

- What I know with confidence: [List]
- What I think I know (needs verification): [List]
- What I know I don't know: [List]

Gate: Scope + definition of done explicit.

Phase 1.5: HYPOTHESIZE

Objective: Transform research into hypothesis testing

1. Generate 3-5 testable hypotheses
2. Assign prior probability: High (70-90%) / Medium (40-70%) / Low (10-40%)
3. Design research to CONFIRM or DISCONFIRM each
4. Track probability shifts as evidence accumulates

Requirements:

• At least 3 hypotheses before Phase 2
• Include at least one contrarian hypothesis
• Track probability shifts in final report

Gate: 3+ hypotheses generated.

Phase 2: PLAN

Objective: Create intelligent research roadmap

1. Identify 5-10 primary sources
2. List 5-10 secondary/backup sources
3. Create 15-25 search query variations using QUEST Matrix
4. Plan triangulation approach

QUEST Matrix: Systematic Query Generation

Use query-framework.md for comprehensive query coverage.

THINK Step: Hunt Direction (H)

Decompose the core question into prioritized sub-questions:

- Factual questions (What is?): [List with priority]
- Causal questions (Why?): [List with priority]
- Evaluative questions (How good?): [List with priority]
- Dependencies: [Which questions must come first?]

Gate: Each subquestion has 3+ queries and 2+ source classes. QUEST checklist ≥80% complete.

Phase 2.5: PLAN PREVIEW

Objective: User validation before executing (Deep+ tier only)

Present plan and ask: "ต้องการปรับแผนก่อนเริ่มไหมคะ?"

Gate: User confirms OR timeout.

Phase 3: RETRIEVE

Objective: Collect information using PARALLEL execution

Parallel Execution (MANDATORY)

[Single message with multiple tool calls]
WebSearch: "[topic] 2025 state of the art"
WebSearch: "[topic] limitations challenges"
WebSearch: "[topic] vs alternatives comparison"
Task(agent): Academic paper analysis

Backtrack Protocol

When: 3+ consecutive searches yield <10% new info

1. STOP current path
2. LOG: "⚠️ Dead-end: [query] → [reason]"
3. PIVOT: Alternative angle/terminology
4. RESUME with new queries

URL Fetching

Primary: WebFetch tool

Fallback (if 403/blocked): Jina AI Reader

curl -s --max-time 60 "https://r.jina.ai/https://example.com/page"

Gate: Each subquestion has ≥3 sources and ≥1 high-quality.

Phase 3.5: GAP ANALYSIS

Objective: Identify gaps, generate refined queries

After initial retrieval, assess:

• ✅ Questions answered
• ⚠️ Questions partially answered
• ❌ Questions unanswered
• Claims needing more evidence

COMPASS Checklist: Perspective Coverage Audit

Use perspective-checklist.md to verify comprehensive coverage.

Min Coverage: Quick 3/7, Standard 4/7, Deep 5/7, Exhaustive 6/7

THINK Step: Inspect Gaps (I)

Reflect on what's still missing:

- Information gathered so far: [Summary]
- Contradictions found: [List with resolution approach]
- Most concerning gap: [What + Why it matters + How to address]
- Follow-up queries: [Specific queries to fill gaps]

Iteration Rules:

• Quick tier: 0-1 iterations
• Standard: 1-2 iterations
• Deep: 2-3 iterations
• Exhaustive: Until diminishing returns

Gate: Gap analysis + COMPASS audit completed; follow-up queries executed.

Phase 4: TRIANGULATE

Objective: Validate information, assign claim types

Claim Taxonomy

Confidence Reasoning (Required for C1)

**Claim:** [Statement]
**Confidence:** HIGH/MEDIUM/LOW/SPECULATIVE
**Reason:** [Why this level]
**What would raise/lower:** [Evidence needed]
**Sources:** [1][2][3]

Confidence Levels

Gate: All C1 claims verified or marked unverified.

Phase 5: SYNTHESIZE

Objective: Generate insights with Implications Engine

For every major finding, answer:

Output format:

**ผลกระทบ (Implications):**
- **แล้วยังไง:** [Significance]
- **แล้วต้องทำอะไร:** [Action]
- **ถ้าเป็นอย่างนี้ต่อ:** [Scenario]

THINK Step: Notice Assumptions (N)

Before drawing conclusions, surface hidden assumptions. See researcher-thinking.md.

- My hidden assumptions: [List with "if wrong" impact]
- Source assumptions: [What my sources assume]
- Missing perspectives: [Whose voice is absent?]

Gate: Every recommendation links to C1/C2 claims. Assumptions surfaced.

Phase 6: RED TEAM

Objective: Find counter-evidence (Devil's Advocate)

Deploy when: Type D investigation OR user requests

Search for:

1. Data contradicting main findings
2. Case studies where approach FAILED
3. Expert opinions that DISAGREE
4. Methodological weaknesses
5. Edge cases where conclusions don't hold

Requirements:

• Present counterarguments at their STRONGEST
• Include "What would change our mind" triggers
• Section title: "ข้อจำกัดและหลักฐานที่ขัดแย้ง"

THINK Step: Know Limits (K)

Practice epistemic humility. See researcher-thinking.md.

- What I could be wrong about: [Top 3 claims + why might be wrong]
- Evidence that would change my mind: [For each conclusion]
- Honest confidence: Overall [HIGH/MEDIUM/LOW]
  - Most certain: [Claim]
  - Least certain: [Claim]

Limitations Statement (for final report): "This analysis has the following limitations: [List 2-3 key limitations and their impact]"

Phase 6.5: SELF-CRITIQUE

Objective: Review quality before packaging

Checklist

• Every C1 has citation + confidence
• Quotes are verbatim
• Evidence supports conclusions
• HIGH claims have 3+ sources
• All hypotheses have outcomes
• Red Team section present

Common Issues

Gate: High-priority issues resolved.

Phase 7: PACKAGE

Objective: Deliver professional research report

Progressive Assembly

1. Executive Summary → Write
2. Hypothesis Results → Append
3. Introduction → Append
4. Findings (each with implications) → Append
5. Red Team / Limitations → Append
6. Recommendations → Append
7. Bibliography (ALL entries) → Append

Anti-Truncation Rules

FORBIDDEN:

• "Content continues..."
• "Due to length..."
• Bibliography ranges "[8-75]"

REQUIRED:

• Complete each section fully
• Every citation [N] has full entry

Quality Checklist

Before delivery:

• Every C1 has citation + confidence + reasoning
• C1 claims have 2+ independent sources
• Contradictions acknowledged
• Sources recent (<3 months for AI/tech)
• No unsupported claims
• Implications for each finding
• Red Team section included
• Self-critique completed
• Gap analysis performed
• Hypothesis outcomes reported