Deep Research

Deep Research

Create high-fidelity research reports with strict format control, evidence mapping, source governance, and multi-pass synthesis.

Architecture: Lead Agent + Subagents

Lead Agent (coordinator — minimizes raw search context)
  |
  P0: Environment + source policy setup
  |
  P1: Research Task Board (roles, queries, parallel groups)
  |
  Dispatch ──→ Subagent A ──→ writes task-a.md ──┐
           ──→ Subagent B ──→ writes task-b.md ──┤ (parallel)
           ──→ Subagent C ──→ writes task-c.md ──┘
  |                                               |
  |     research-notes/  <────────────────────────┘
  |
  P2: Build citation registry with source_type + as_of + authority
  P3: Evidence-mapped outline with counter-claim flags
  P4: Draft from notes (never from raw search results)
  P5: Counter-review (claims, confidence, alternatives)
  P6: Verify (every [n] in registry, traceability check)
  P7: Polish → final report with confidence markers

Context efficiency: Subagents' raw search results stay in their context and are discarded. Lead agent sees only distilled notes (~60-70% context reduction).

Mode Selection

Determine the research mode before starting:

• Enterprise Research Mode: Six-dimension data collection with structured analysis frameworks (SWOT, risk matrix, competitive barrier quantification)
• General Research Mode: Standard P0-P7 research pipeline with source governance
• Depth Selection: Lightweight for single entity/concept < 30 words; Standard for multi-entity comparison or "深入"/"comprehensive" requests

Source Governance (V6)

Source Accessibility Classification

CRITICAL RULE: Every source must be classified by accessibility:

⚠️ CIRCULAR VERIFICATION BAN: You must NOT:

• Use user's private data to "discover" what they already know about themselves
• Research user's own company by accessing user's private accounts
• Present user's private knowledge as "research findings"

✅ EXCLUSIVE INFORMATION ADVANTAGE: You SHOULD:

• Use user's Crunchbase Pro to research competitors
• Use user's proprietary databases for market research
• Use user's private APIs for investment analysis
• Leverage any exclusive source user provides for third-party research

Source Type Labels

Every source MUST also be tagged with:

Quality Gates:

• Standard mode: ≥30% official sources in final approved set
• Lightweight mode: ≥20% official sources
• Maximum single-source share: ≤25% (Standard), ≤30% (Lightweight)
• Minimum unique domains: 5 (Standard), 3 (Lightweight)

AS_OF Date Policy

Set AS_OF date explicitly at P0. For all time-sensitive claims:

• Include source publication date with every citation
• Downgrade confidence if source is older than relevant horizon
• Flag stale sources in registry (studies >3 years, news >6 months for fast-moving topics)

P0: Environment & Policy Setup

Check capabilities before starting:

Set policy variables:

• AS_OF: Today's date (YYYY-MM-DD) - mandatory for timed topics
• MODE: Standard (default) or Lightweight
• SOURCE_TYPE_POLICY: Enforce official/academic/secondary/journalism/community/other labels
• COUNTER_REVIEW_PLAN: What opposing interpretation to test

Report: [P0 complete] Subagent: {yes/no}. Mode: {standard/lightweight}. AS_OF: {YYYY-MM-DD}.

When researching a specific company/enterprise, follow this specialized workflow that ensures six-dimension coverage, quantified analysis frameworks, and three-level quality control.

Enterprise Workflow Overview

Enterprise Research Progress:
- [ ] E1: Intake — confirm company entity, research depth, format contract
- [ ] E2: Six-dimension data collection (parallel where possible)
  - [ ] D1: Company fundamentals (entity, founding, funding, ownership)
  - [ ] D2: Business & products (segments, products, revenue structure)
  - [ ] D3: Competitive position (industry rank, competitors, barriers)
  - [ ] D4: Financial & operations (3-year financials, efficiency metrics)
  - [ ] D5: Recent developments (6-month events, strategic signals)
  - [ ] D6: Internal/proprietary sources (or note limitation)
- [ ] E3: Structured analysis frameworks
  - [ ] SWOT analysis (evidence-backed, 4 quadrants × 3-5 entries)
  - [ ] Competitive barrier quantification (7 dimensions, weighted score)
  - [ ] Risk matrix (8 categories, probability × impact)
  - [ ] Comprehensive scorecard (6 dimensions, weighted total)
- [ ] E4: L1/L2/L3 quality checks at each stage transition
- [ ] E5: Draft report using 7-chapter enterprise template
- [ ] E6: Multi-pass drafting + UNION merge (same as general Step 6-7)
- [ ] E7: Present draft for human review and iterate

P1: Research Task Board

Decompose the research question into 4-6 investigation tasks (Standard) or 3-4 tasks (Lightweight).

Each task assignment includes:

• Expert Role: Specialist persona (e.g., "Policy Historian", "Ecosystem Mapper")
• Objective: One-sentence investigation goal
• Queries: 2-3 pre-planned search queries
• Depth: DEEP (fetch 2-3 full articles) or SCAN (snippets sufficient)
• Output: Path to research notes file
• Parallel Group: Group A (independent) or Group B (depends on Group A)

Task Decomposition Rules

1. Each task covers one coherent sub-topic a specialist would own 2. Group A tasks must be independent and source-diverse 3. Max 3 tasks per parallel group (concurrency limit) 4. Every task must flag time-sensitive claims and expected citation aging risk

Enterprise Research Integration

When in Enterprise Research Mode, task board maps to six dimensions:

• Task A: Company fundamentals (entity, founding, funding, ownership)
• Task B: Business & products (segments, products, revenue structure)
• Task C: Competitive position (industry rank, competitors, barriers)
• Task D: Financial & operations (3-year financials, efficiency metrics)
• Task E: Recent developments (6-month events, strategic signals)
• Task F: Internal/proprietary sources (or document limitation)

Report: [P1 complete] {N} tasks in {M} groups. Dispatching Group A.

---

Enterprise Research Mode (Specialized Pipeline)

When researching a specific company/enterprise, follow this specialized workflow that ensures six-dimension coverage, quantified analysis frameworks, and three-level quality control.

E1: Intake

Same as P0/P1 above, plus:

• Confirm the exact legal entity being researched (parent vs subsidiary)
• Select research depth: Quick scan (3-5 pages) / Standard (10-20 pages) / Deep (20-40 pages)
• Identify any specific comparison targets (benchmark companies)

P2: Dispatch + Investigate

Subagents execute tasks using references/subagent_prompt.md and output to references/research_notes_format.md.

With Subagents (Claude Code / Cowork / DeerFlow)

1. Dispatch Group A tasks in parallel (max 3 concurrent) 2. Each subagent searches, fetches, and tags source types 3. Every source line includes Source-Type and As Of 4. Wait for Group A completion 5. Dispatch Group B (can read Group A notes)

Subagent Output Requirements

Each task-{id}.md must contain:

• Sources section: URLs from actual search results with Source-Type, As Of, Authority (1-10)
• Findings section: Max 10 one-sentence facts with source numbers
• Deep Read Notes (DEEP tasks): 2-3 sources read in full with key data/insights
• Gaps section: What was searched but NOT found, alternative interpretations

Without Subagents (Degraded Mode)

Lead agent executes tasks sequentially, acting as each specialist. Raw search results are discarded after writing notes.

Enterprise Research: Six-Dimension Collection

Follow references/enterprise_research_methodology.md for:

• Detailed collection workflow per dimension (query strategies, data fields, validation)
• Data source priority matrix (P0-P3 ranking)
• Cross-validation rules (min sources, max deviation thresholds)

Key principles:

• Evidence-driven: every conclusion must trace to a citable source
• Multi-source validation: key data requires ≥2 independent sources
• Restrained judgment: mark speculation explicitly, avoid unsubstantiated claims
• Structured presentation: complex information via tables, lists, hierarchies

Run L1 quality check after completing each dimension (see enterprise_quality_checklist.md).

Status per task: [P2 task-{id} complete] {N} sources, {M} findings. Status all: [P2 complete] {N} tasks done, {M} total sources. Building registry.

E3: Structured Analysis Frameworks

Apply frameworks from references/enterprise_analysis_frameworks.md in order: 1. SWOT analysis — each entry with evidence + source + impact assessment 2. Competitive barrier quantification — 7 dimensions with weighted scoring → A+/A/B+/B/C+/C rating 3. Risk matrix — 8 mandatory categories, probability × impact → Red/Yellow/Green 4. Comprehensive scorecard — 6-dimension weighted total → X/10

Run L2 quality check after analysis is complete.

E4: Quality Control

Three-level checks from references/enterprise_quality_checklist.md:

• L1 (Data): Source count, attribution, cross-validation, timeliness
• L2 (Analysis): SWOT completeness, risk coverage, barrier scoring, conclusion support
• L3 (Document): Structure compliance, format consistency, readability, appendices

E5: Draft Using Enterprise Template

Use the 7-chapter enterprise report template from enterprise_quality_checklist.md: 1. Company Overview 2. Business & Product Structure 3. Market & Competitive Position 4. Financial & Operations Analysis 5. Risks & Concerns 6. Recent Developments 7. Comprehensive Assessment & Conclusion

Plus appendices: Data Source Index, Glossary, Disclaimer.

E3-E7: Enterprise Analysis, Drafting, and Review

• E3: Structured Analysis — Apply frameworks from references/enterprise_analysis_frameworks.md
• E4: Quality Control — Run L1/L2/L3 checks per references/enterprise_quality_checklist.md
• E5: Draft — Use 7-chapter enterprise template
• E6-E7: Multi-Pass Drafting and Review — Same as P4-P7 below

---

P3: Citation Registry + Source Governance

Lead agent reads all task notes and builds unified registry.

Registry Process

1. Read every task file's ## Sources section 2. Merge all sources, deduplicate by URL 3. Assign sequential [n] numbers by first appearance 4. Tag: source_type, as_of date, authority score (1-10), task id 5. Apply quality gates:

• Standard: ≥12 approved sources, ≥5 unique domains, ≥30% official
• Lightweight: ≥6 approved sources, ≥3 unique domains, ≥20% official
• Max single-source share: ≤25% (Standard), ≤30% (Lightweight)

6. Drop sources below threshold and list them explicitly

Registry Output Format

CITATION REGISTRY

Approved:
[1] Author/Org — Title | URL | Source-Type: official | Accessibility: public | Date: 2026-03-01 | Auth: 8 | task-a
[2] ...

Dropped:
x Source | URL | Source-Type: community | Accessibility: privileged | Auth: 3 | Reason: PRIVILEGED SOURCE - NOT ALLOWED

Stats: {approved}/{total}, {N} domains, official_share {xx}%
Privileged sources rejected: {N}

Critical rule: These [n] are FINAL. P5 may only cite from Approved list. Dropped sources never reappear.

Circular verification handling: When researching the user's own company/assets, if you discover data in user's private accounts (e.g., user's domain registrar showing they own domains), you MUST: 1. Reject it from the registry (user already knows this) 2. Note it as "CIRCULAR - USER ALREADY KNOWS" in Dropped 3. Search for equivalent PUBLIC sources (e.g., public WHOIS, news articles) 4. Report from external investigator perspective only

Exclusive source handling: When user EXPLICITLY PROVIDES their paid subscriptions or private APIs for third-party research (e.g., "Use my Crunchbase Pro to research competitors"), you SHOULD: 1. Accept it as "exclusive-user-provided" accessibility 2. Use it as competitive advantage 3. Cite it properly in registry 4. If no public equivalent exists, mark as [unverified] or omit the claim

Report: [P3 complete] {approved}/{total} sources. {N} domains. Official share: {xx}%. Privileged rejected: {N}.

Handling Information Black Box

When researching entities with no public footprint (like the "字节跳动子公司" example):

What an external researcher would find:

• WHOIS: Privacy protected → No owner info
• Web search: No news, no press releases
• Social media: No company pages
• Business registries: No public API or requires local access
• Result: Complete information black box

Correct response:

Findings: NO PUBLIC INFORMATION AVAILABLE

Sources checked:
- WHOIS (public): Privacy protected [failed]
- Company registry (public): Access denied/No API [failed]
- News media: No coverage [failed]
- Corporate website: Placeholder only [minimal]

Verdict: UNABLE TO VERIFY COMPANY EXISTENCE from external perspective
Sources found: 0 (or minimal, e.g., only WHOIS showing domain exists)
Confidence: N/A - Insufficient evidence

DO NOT:

• ❌ Use user's own credentials to "fill in the gaps"
• ❌ Assume the company exists based on domain registration alone
• ❌ Fill missing data with speculation
• ❌ Claim to have "verified" information you accessed through privileged means

DO:

• ✅ Clearly state what an external researcher can/cannot verify
• ✅ Document all failed search attempts
• ✅ Mark claims as [unverified] or omit entirely
• ✅ Downgrade mode to Lightweight or stop if insufficient public sources
• ✅ Recommend direct contact for due diligence

---

P4: Evidence-Mapped Outline

Lead agent reads notes + registry to build outline.

1. Identify cross-task patterns 2. Design sections topic-first, not task-order-first 3. Map each section to specific findings with source numbers 4. Flag sections needing counter-review 5. Mark recency-sensitive claims with AS_OF checks

Outline format:

## N. {Section Title}
Sources: [1][3][7] from tasks a, b
Claims: {claim from task-a finding 3}, {claim from task-b finding 1}
Counter-claim candidates: {alternative explanations}
Recency checks: {source dates + AS_OF}
Gaps: {limited official evidence}

---

P5: Draft from Notes

Write section by section using references/report_template_v6.md.

Rules:

• Every factual claim needs citation [n]
• Numbers/percentages must have source
• Add confidence marker per section: High/Medium/Low with rationale
• Add counter-claim sentence when evidence conflicts
• No new sources may be introduced
• Use [unverified] for unsupported statements

Anti-hallucination:

• Lead agent never invents URLs — only from subagent notes
• Lead agent never fabricates data — mark [unverified] if number not in notes

Status: [P5 in progress] {N}/{M} sections, ~{words} words.

---

P6: Counter-Review (Mandatory)

For each major conclusion, perform opposite-view checks:

1. Could the conclusion be wrong? 2. Which high-impact claims depend on a single source? 3. Which claims lack official/academic support? 4. Are stale sources used for time-sensitive claims? 5. Find ≥3 issues (re-examine if 0 found)

Using Counter-Review Team (Recommended)

For comprehensive parallel review, use the Counter-Review Team:

# 1. Prepare inputs
counter-review-inputs/
  ├── draft_report.md
  ├── citation_registry.md
  ├── task-notes/
  └── p0_config.md

# 2. Dispatch to 4 specialist agents in parallel
SendMessage to: claim-validator
SendMessage to: source-diversity-checker
SendMessage to: recency-validator
SendMessage to: contradiction-finder

# 3. Wait for all specialists to complete

# 4. Send to coordinator for synthesis
SendMessage to: counter-review-coordinator
  inputs: [4 specialist reports]

# 5. Receive final P6 Counter-Review Report

See references/counter_review_team_guide.md for detailed usage.

Manual Counter-Review (Fallback)

If Counter-Review Team is unavailable, perform manual checks:

• Verify every high-confidence claim has ≥2 sources
• Check official/academic backing for key claims
• Verify AS_OF dates on time-sensitive claims
• Document opposing interpretations

Output

Include in final report:

## 核心争议 / Key Controversies
- **争议 1:** [主张 A 与反向证据 B 对比] [n][m]
- **争议 2:** ...

Report: [P6 complete] {N} issues found: {critical} critical, {high} high, {medium} medium.

---

P7: Verify

Cross-check before finalization:

1. Registry cross-check: List every [n] in report vs approved registry 2. Spot-check 5+ claims: Trace to task notes 3. Remove/fix non-traceable claims 4. Validate no dropped source resurrected 5. Check source concentration for key claims

Report: [P7 complete] {N} spot-checks, {M} violations fixed.

---

Output Requirements

• Match the requested language and tone
• Preserve technical terms in English
• Respect the report spec and formatting rules
• Include a references section or bibliography

Reference Files

Core V6 Pipeline References

General Research References

Enterprise Research References (load when in Enterprise Research Mode)

Anti-Patterns

• Single-pass drafting without parallel complete passes
• Splitting passes by section instead of full report drafts
• Ignoring the format contract or user template
• Claims without citations or evidence table mapping
• Mixing conflicting dates without calling out discrepancies
• Copying external AI output without verification
• Deleting intermediate drafts or raw research outputs
• Lead agent reading raw search results — only read subagent notes
• Inventing URLs — only use URLs from actual search results
• Resurrecting dropped sources — dropped in P3 never reappear
• Missing AS_OF for time-sensitive claims — always include source date
• Skipping counter-review — mandatory P6 must find ≥3 issues
• CIRCULAR VERIFICATION — never use user's private data to "discover" what they already know about themselves
• IGNORING EXCLUSIVE SOURCES — when user provides Crunchbase Pro etc. for competitor research, USE IT

Next Step: Verify and Deliver

After completing research, suggest verification and output:

Research report complete: [N] sources cited, [M] claims made.

Options:
A) Verify facts — run /fact-checker on the report (Recommended)
B) Create slides — run /daymade-docs:ppt-creator from the findings
C) Export as PDF — run /daymade-docs:pdf-creator for formal delivery
D) No thanks — the report is ready as-is

Completeness Review Checklist

Verify the draft meets all requirements before delivery.

Structure and Format

• All required sections present and ordered
• Headings match the format contract
• Length targets met per section

Evidence and Citations

• Every claim has a citation
• Every numeric claim has at least one Tier A or B source
• Conflicting sources are explicitly noted
• Evidence table maps claims to sources

Content Quality

• Findings answer the research questions
• Recommendations are tied to evidence
• Limitations and uncertainty are documented
• Terminology is consistent

Final Checks

• Dates and time ranges are consistent
• No unsupported claims
• No duplicate or conflicting statements

Counter-Review Team 使用指南

Deep Research V6 P6 阶段的专用 Agent Team，并行执行多维度审查。

Team 架构

counter-review-coordinator (协调者)
    ├── claim-validator (声明验证器)
    ├── source-diversity-checker (来源多样性检查器)
    ├── recency-validator (时效性验证器)
    └── contradiction-finder (矛盾发现器)

Agent 职责

使用流程

1. 准备输入材料

在 P5 (Draft) 完成后，收集以下材料：

inputs/
├── draft_report.md          # P5 起草的报告
├── citation_registry.md     # P3 的引用注册表
├── task-notes/
│   ├── task-a.md           # 子代理研究笔记
│   ├── task-b.md
│   └── ...
└── p0_config.md            # P0 配置 (AS_OF 日期, Mode 等)

2. 并行分发任务

向 4 个 specialist agent 同时发送任务：

# 向 claim-validator 发送
SendMessage to: claim-validator
  输入: draft_report.md + citation_registry.md + task-notes/
  指令: 验证所有声明的证据支持

# 向 source-diversity-checker 发送
SendMessage to: source-diversity-checker
  输入: draft_report.md + citation_registry.md
  指令: 检查来源多样性和单一来源依赖

# 向 recency-validator 发送
SendMessage to: recency-validator
  输入: draft_report.md + citation_registry.md + p0_config.md
  指令: 验证时敏声明的新鲜度

# 向 contradiction-finder 发送
SendMessage to: contradiction-finder
  输入: draft_report.md + task-notes/ + citation_registry.md
  指令: 发现矛盾和缺失的反向观点

3. 协调汇总

等待 4 个 specialist 完成后，发送给 coordinator：

SendMessage to: counter-review-coordinator
  输入:
    - Claim Validation Report
    - Source Diversity Report
    - Recency Validation Report
    - Contradiction and Bias Report
  指令: 整合所有报告，生成最终 P6 Counter-Review Report

4. 获取最终输出

Coordinator 输出包含：

• 问题汇总（必须 ≥3 个）
• 关键争议部分（可直接复制到最终报告）
• 强制修复清单
• 质量门状态

质量门要求

输出示例

Coordinator 最终报告结构

# P6 Counter-Review Report

## Executive Summary
- Total issues found: 7 (critical: 2, high: 3, medium: 2)
- Must-fix before publish: 2
- Recommended improvements: 5

## Critical Issues (Block Publish)
| # | Issue | Location | Source | Fix Required |
|---|-------|----------|--------|--------------|
| 1 | 市场份额声明无来源 | 3.2节 | 无 | 补充来源或删除 |
| 2 | 单一社区来源支持收入数据 | 4.1节 | [12] community | 找官方来源替代 |

## 核心争议 / Key Controversies

- **争议 1:** 公司声称增长 50% vs 分析师报告增长 30%
  - 证据强度: official(公司财报) vs academic(第三方研究)
  - 建议: 并列呈现两种数据，说明差异原因

## Mandatory Fixes Checklist
- [ ] 补充 3.2 节市场份额来源
- [ ] 替换 4.1 节收入数据来源
- [ ] 添加 AS_OF: 2026-04-03 到所有时敏声明

## Quality Gates Status
| Gate | Status | Notes |
|------|--------|-------|
| P6 ≥3 issues found | ✅ | 发现 7 个问题 |
| No critical claim single-sourced | ❌ | 2 个问题待修复 |
| AS_OF dates present | ❌ | 3 处缺失 |
| Counter-claims documented | ✅ | 已添加 |

集成到 SKILL.md 工作流

在 SKILL.md 的 P6 阶段，添加以下指令：

## P6: Counter-Review (Mandatory)

**使用 Counter-Review Team 执行并行审查：**

1. **准备材料**: draft_report.md, citation_registry.md, task-notes/, p0_config.md
2. **并行分发**: 同时发送给 4 个 specialist agent
3. **等待完成**: 收集 4 份 specialist 报告
4. **协调汇总**: 发送给 coordinator 生成最终 P6 报告
5. **强制执行**: 所有 Critical 问题必须在 P7 前修复
6. **输出**: 将"核心争议"部分复制到最终报告

**Report**: `[P6 complete] {N} issues found: {critical} critical, {high} high, {medium} medium.`

团队管理

查看团队状态

cat ~/.claude/teams/counter-review-team/config.json

向 Agent 发送消息

SendMessage to: claim-validator
message: 开始审查任务，输入文件在 ./review-inputs/

关闭团队

SendMessage to: "*"
message: {"type": "shutdown_request", "reason": "任务完成"}

注意事项

1. 必须发现 ≥3 个问题 - 如果 coordinator 报告 <3 个问题，需要重新审查 2. Critical 问题必须修复 - 才能进入 P7 3. 保留所有审查记录 - 作为研究方法论的一部分 4. 中文输入中文输出 - 所有 agent 支持中英文双语

Enterprise Analysis Frameworks

Apply these frameworks after completing the six-dimension data collection. Execute in order: SWOT → Competitive Barriers → Risk Matrix → Comprehensive Scoring.

SWOT Analysis Template

Each SWOT entry MUST include evidence and source attribution.

|              | Positive Factors                  | Negative Factors                  |
|--------------|-----------------------------------|-----------------------------------|
| **Internal** | **S (Strengths)**                 | **W (Weaknesses)**                |
|              | 1. {description}                  | 1. {description}                  |
|              |    • Evidence: {data/fact}        |    • Evidence: {data/fact}        |
|              |    • Source: {citation}           |    • Source: {citation}           |
|              |    • Impact: {assessment}         |    • Impact: {assessment}         |
|              |                                   |                                   |
| **External** | **O (Opportunities)**             | **T (Threats)**                   |
|              | 1. {description}                  | 1. {description}                  |
|              |    • Evidence: {trend/policy}     |    • Evidence: {pressure/risk}    |
|              |    • Source: {citation}           |    • Source: {citation}           |
|              |    • Probability: {assessment}    |    • Probability: {assessment}    |
|              |    • Impact: {assessment}         |    • Impact: {assessment}         |

Requirements:

• Each quadrant: 3-5 entries minimum
• Every entry must have evidence with source
• S/W must be data-backed (not opinions)
• O/T must include probability and impact estimates

Strategic Implications Matrix (generate after SWOT):

• SO Strategy (leverage strengths to capture opportunities): 1-2 specific recommendations
• WO Strategy (overcome weaknesses to seize opportunities): 1-2 specific recommendations
• ST Strategy (use strengths to counter threats): 1-2 specific recommendations
• WT Strategy (mitigate weaknesses to avoid threats): 1-2 specific recommendations

Competitive Barrier Quantification Framework

7 barrier dimensions with weighted scoring:

Scoring: Total = Σ(dimension score × weight)

Rating Scale:

Output format: Present a scorecard table with each dimension's strength rating, raw score, justification (with evidence), and the weighted total with final rating.

Risk Matrix Framework

Assess 8 mandatory risk categories:

Risk Assessment Scales

Probability:

Impact:

Risk Level: Risk Value = Probability Score × Impact Score

8 Mandatory Risk Categories

Risk Table Format

Requirements:

• All 8 categories must be assessed (no skipping)
• Each risk entry must cite specific evidence or triggers
• Provide current mitigations AND recommended actions
• High risks: require immediate action plans
• Medium risks: require monitoring plans
• Low risks: require periodic review schedule

Comprehensive Scoring (Final Section)

After completing SWOT, barriers, and risk matrix, generate a comprehensive scorecard:

| Dimension | Score | Weight | Weighted | Key Evidence |
|-----------|-------|--------|----------|-------------|
| Business Quality | X/10 | 25% | | |
| Competitive Position | X/10 | 20% | | |
| Financial Health | X/10 | 20% | | |
| Growth Potential | X/10 | 15% | | |
| Risk Profile | X/10 | 10% | | |
| Management Quality | X/10 | 10% | | |
| **Total** | | 100% | **X/10** | |

Every score must reference specific evidence from the six-dimension data collection.

Enterprise Research Quality Checklist

Three-level quality control executed at each stage transition.

L1: Data Collection Quality (after each dimension)

Per-Dimension Checks

Result handling: All pass → proceed. Partial fail → supplement sources. Critical fail → re-collect dimension.

Dimension-Specific Checklists

D1 Company Fundamentals (target: 11/11):

• [ ] Legal entity boundaries clarified
• [ ] Founding date with month/year
• [ ] Headquarters city identified
• [ ] Founder/CEO confirmed (≥2 sources)
• [ ] Employee count with year
• [ ] Listing status (exchange, ticker)
• [ ] Latest valuation/market cap with date
• [ ] Core business one-liner
• [ ] Funding history ≥3 rounds
• [ ] ≥5 milestone events in timeline
• [ ] Ownership structure: controller identified

D2 Business & Products (target: 7/7):

• [ ] ≥3 business segments identified
• [ ] Revenue share per segment
• [ ] ≥3 core products analyzed
• [ ] User metrics (DAU/MAU) with numbers
• [ ] Monetization model per product
• [ ] Revenue breakdown (segment/geography/customer)
• [ ] Growth/decline trend per segment

D3 Competitive Position (target: 7/7):

• [ ] Industry clearly defined
• [ ] Market size quantified
• [ ] Company rank established
• [ ] Market share with number
• [ ] ≥3 competitors identified
• [ ] Multi-dimension comparison table complete
• [ ] ≥5 barrier dimensions assessed with scores

D4 Financial & Operations (target: 9/9):

• [ ] Revenue: 3-year data
• [ ] Net income: 3-year data
• [ ] Gross margin: 3-year data
• [ ] Net margin: 3-year data
• [ ] Operating cash flow: 3-year data
• [ ] R&D expense: 3-year data
• [ ] Key financial data cross-validated (≥2 sources)
• [ ] Metric definitions consistent across years
• [ ] ≥3 efficiency metrics (ROE/ROA/etc.)

D5 Recent Developments (target: 5/5):

• [ ] ≥5 recent events (within 6 months)
• [ ] Events span ≥3 event types
• [ ] Each event has impact assessment
• [ ] ≥2 strategic direction signals identified
• [ ] Most recent event within 1 month

D6 Internal/Proprietary (target: 2/2):

• [ ] Internal knowledge base queried (or limitation noted)
• [ ] Internal document search executed (or limitation noted)

L2: Analysis Quality (after analysis frameworks applied)

Result handling: All pass → proceed to writing. Partial fail → supplement analysis evidence. Critical fail → re-execute analysis framework.

L3: Document Quality (after report drafted)

Result handling: All pass → deliver. Partial fail → format optimization. Critical fail → regenerate document.

Enterprise Report Structure (7 Chapters)

# {Company Name} Research Report

> Executive Summary: {1-2 sentence core conclusion}

---

## 1. Company Overview
### 1.1 Basic Information (table)
### 1.2 Development Timeline
### 1.3 Funding History (table)
### 1.4 Ownership Structure & Control
### 1.5 Core Management Team (table)

## 2. Business & Product Structure
### 2.1 Business Landscape Overview
### 2.2 Core Product Matrix (table)
### 2.3 Revenue Structure Analysis
### 2.4 Business Development Trends

## 3. Market & Competitive Position
### 3.1 Industry Position Analysis
### 3.2 Competitive Comparison (table)
### 3.3 Competitive Barrier Assessment (scorecard)

## 4. Financial & Operations Analysis
### 4.1 Key Financial Metrics (3-year comparison table)
### 4.2 Operating Efficiency Assessment
### 4.3 Financial Health Summary

## 5. Risks & Concerns
### 5.1 Risk Matrix Analysis (8-category table)
### 5.2 Key Risk Deep-Dives
### 5.3 Risk Mitigation Recommendations

## 6. Recent Developments
### 6.1 Major Recent Events (table)
### 6.2 Strategic Signal Interpretation

## 7. Comprehensive Assessment & Conclusion
### 7.1 SWOT Summary
### 7.2 Comprehensive Scorecard
### 7.3 Core Conclusions & Outlook

---

## Appendices
### A. Data Source Index
### B. Glossary
### C. Disclaimer

Quality Control Four Dimensions

Apply throughout all stages:

Enterprise Research Methodology

Six-Dimension Data Collection

Enterprise research requires parallel collection across six dimensions. Execute all six in order, writing findings to a structured draft after each dimension.

Dimension 1: Company Fundamentals

Step 1.1: Confirm legal entity
├── Clarify parent/subsidiary/affiliate boundaries
├── Query: "{company} legal entity corporate structure"
├── Output: Entity scope statement
└── Verify: Map operating entities to brands

Step 1.2: Basic information
├── Query round 1: "{company} founding date headquarters founder"
├── Query round 2: "{company} company overview profile"
├── Query round 3: "{company} CEO management team executives"
├── Source priority: Official site > Regulatory filings > Authoritative media
└── Output: Basic info table (name, founded, HQ, CEO, employees, listing status)

Step 1.3: Funding history
├── Query: "{company} funding rounds valuation IPO"
├── Key fields: round, amount, investors, post-money valuation, date
└── Output: Funding timeline table

Step 1.4: Ownership structure
├── Query: "{company} ownership structure beneficial owner"
├── Key fields: controller identity, economic interest %, voting rights %, control mechanisms (dual-class etc.)
└── Output: Ownership summary

Dimension 2: Business & Products

Step 2.1: Business landscape scan
├── Query round 1: "{company} product lines business segments"
├── Query round 2: "{company} revenue breakdown by segment"
├── Query round 3: "{company} business model monetization"
├── Key fields: segment name, positioning, revenue share, YoY growth, synergies
└── Output: Business landscape table

Step 2.2: Core product analysis
├── Query: "{company} core products DAU MAU user base"
├── Per product: positioning, target users, scale (DAU/MAU), market share, monetization, competitive advantage, trends
└── Output: Product matrix table

Step 2.3: Revenue structure analysis
├── Source: Financial reports (deep extraction)
├── Breakdown by: segment, geography, customer type, pricing model
└── Output: Revenue structure summary

Dimension 3: Competitive Position

Step 3.1: Industry position
├── Query: "{company} industry ranking market share"
├── Key fields: industry definition, TAM/SAM/SOM, company rank, share, concentration (CR3/CR5)
└── Output: Industry position analysis

Step 3.2: Competitor identification & comparison
├── Query round 1: "{company} competitors"
├── Query round 2: "{company} vs {competitor A} comparison"
├── Query round 3: "{company} vs {competitor B} differences"
├── Comparison dimensions: founding, revenue, market share, core products, user scale, valuation/market cap, strengths, weaknesses
├── Minimum: ≥3 competitors identified
└── Output: Competitive comparison table

Step 3.3: Competitive barriers assessment
├── Use quantified barrier framework (see enterprise_analysis_frameworks.md)
├── 7 dimensions: network effects, scale economies, brand, technology/patents, switching costs, regulatory licenses, data assets
└── Output: Barrier scorecard with rating

Dimension 4: Financial & Operations

Step 4.1: Financial data collection
├── Query: "{company} financial results {year} revenue profit"
├── Core metrics (3-year minimum): revenue, revenue growth, net income, gross margin, net margin, operating cash flow, R&D expense, R&D ratio
└── Output: Financial metrics table (3+ years)

Step 4.2: Operating efficiency analysis
├── Query: "{company} ROE ROA efficiency per-employee"
├── Efficiency metrics: ROE, ROA, revenue per employee, accounts receivable days, debt-to-equity
└── Output: Operating efficiency table

Step 4.3: Cross-validation
├── Require ≥2 independent sources for key financial data
├── Sources: company filings (primary), regulatory filings, authoritative financial data providers
├── Deviation rules:
│   ├── ≤10%: Pass
│   ├── 10-20%: Flag with explanation
│   └── >20%: Require third-party verification
└── Output: Validation record

Dimension 5: Recent Developments

Step 5.1: Recent news scan (past 6 months)
├── Query round 1: "{company} latest news {current year}"
├── Query round 2: "{company} strategy pivot latest developments"
├── Query round 3: "{company} executive changes leadership"
├── Query round 4: "{company} partnership acquisition latest"
├── Query round 5: "{company} product launch new release"
├── Event types: product launches, fundraising/capital, strategy shifts, executive changes, M&A/partnerships, regulatory/compliance
├── Minimum: ≥5 events identified
└── Output: Major events table

Step 5.2: Strategic signal interpretation
├── Dimensions: expansion signals, contraction signals, transformation signals, risk signals
└── Output: Strategic signal analysis

Dimension 6: Internal/Proprietary Sources

Step 6.1: Internal knowledge base query (if available)
├── Query 1: "our company's relationship with {target company}"
├── Query 2: "internal assessment of {target company}"
├── Query 3: "{target company} competitive analysis"
├── Query 4: "{target company} industry research"
└── Output: Internal perspective supplementary info

Step 6.2: If no internal sources available
├── State explicitly: "No internal/proprietary sources available for this research"
├── Compensate with additional public source depth
└── Note limitation in final report

Data Source Priority Matrix

Rule: P0 + P1 are primary sources. P2 for validation. P3 for reference only, never as sole source.

Cross-Validation Rules

Search Strategy Best Practices

1. Multi-angle queries: 3 different query angles per topic 2. Time filtering: Prioritize data within last 12 months for operational data, last 3 years for financial trends 3. Site restriction: Use site: for authoritative domains when possible 4. Language diversity: Query in both English and the company's primary language 5. Exclude noise: Use - to exclude irrelevant results 6. Progressive depth: Start broad, then narrow based on gaps identified

Formatting Rules

Use these rules to enforce strict report formatting.

Headings

• Use H2 for top-level sections
• Use H3 for subsections
• Keep heading titles consistent with the report spec

Section Order

• Follow the report spec order exactly
• Do not add or remove sections without approval

Bullets and Tables

• Use bullets for lists of findings and actions
• Use tables for comparisons, metrics, or timelines
• Keep tables compact and label all columns

Citations

• Use the citation style defined in the report spec
• Place citations immediately after the claim
• If a claim has multiple sources, list all relevant sources

Terminology

• Define key terms once and reuse consistently
• Preserve technical terms in English

Formatting Hygiene

• Avoid mixed numbering styles in the same section
• Avoid inline URLs in prose unless the spec requires it
• Do not embed new assets unless requested

Quality Gates V6

Gate 1: Task Notes Quality (after P2)

Gate 2: Citation Registry (after P3)

Gate 3: Draft Quality (after P5)

Gate 4: Notes Traceability (after P6)

Gate 5: Verification (after P7)

Anti-Hallucination Patterns

Chinese-Specific Patterns

{{TITLE}}

研究日期: {{DATE}} | 来源数量: {{SOURCE_COUNT}} | 字数: ~{{WORD_COUNT}} | 模式: {{MODE}} | AS_OF: {{AS_OF}} | 官方源占比: {{OFFICIAL_SHARE}}

摘要 / Executive Summary

{{200-400 words summarizing key findings, methodology, conclusions, and risks.}}

---

目录

{{Auto-generate from actual section headers below.}}

---

{{BODY SECTIONS — Adapt to topic type and include opposing interpretation per section.}}

For each section:

N. [Topic-Specific Section Title]

{{Section content with inline citations [1][2]. Standard mode: 500-1000 words per section. Lightweight mode: 300-600 words per section.

Rules:

• 每个事实性论点都需要引用 [n]
• 数字/百分比必须有来源
• 出现不同证据时要成对给出支持与反驳

}}

置信度: High/Medium/Low

依据: {{Why this confidence level — source agreement, evidence quality, data availability}}

反方解释: {{One explicit opposing interpretation with supporting citations if any, or [unverified] if insufficient.}}

---

{{COUNTER-REVIEW SUMMARY}}

• 核心争议 1: [主张 A 与反向证据 B 对比] [n][m]
• 核心争议 2: ...

关键发现 / Key Findings

{{3-5 findings in Standard mode, 2-3 in Lightweight. Each finding should:}}

• 具体结论
• 对应引文
• 信心说明

Example:

• 发现 1: [Most important discovery] [3][7]
• 发现 2: [Second most important] [1][4]

---

局限性与未来方向 / Limitations & Future Directions

本研究局限

{{Be explicit:

• What topics/angles couldn't be covered and why
• Methodological limits (web-accessible sources, paywall, language, timing)
• Source coverage gaps and counter-claim evidence gaps

}}

未来方向

{{Concrete suggestions for follow-up research with priority and responsible evidence type.}}

---

参考文献 / References

[1] Author/Org. "Title". Source-Type: official/academic/secondary-industry/journalism/community/other. As Of: YYYY-MM-DD. URL. [2] Author/Org. "Title". Source-Type: ... As Of: YYYY-MM-DD. URL.

Rules:

• Every [n] in body MUST have matching entry here
• Every entry here MUST be cited at least once
• Source-Type and As Of fields are mandatory
• All URLs MUST come from actual search results (P2 source pool)

Research Notes Format Specification

The research notes are the ONLY communication channel between subagents and the lead agent. Every fact in the final report must be traceable to a line in these notes. No exceptions.

File Structure

workspace/research-notes/
  task-a.md       Subagent A writes (history expert)
  task-b.md       Subagent B writes (transport historian)
  task-c.md       Subagent C writes (telecom analyst)
  task-d.md       Subagent D writes (comparative analyst)
  registry.md     Lead agent builds from task-*.md (P3)

Per-Task Notes Format

Each task-{id}.md file follows this exact structure:

---
task_id: a
role: Economic Historian
status: complete
sources_found: 4
---

## Sources

[1] Before AI skeptics, Luddites raged against the machine | https://www.nationalgeographic.com/... | Source-Type: secondary-industry | As Of: 2025-08 | Authority: 8/10
[2] Rage against the machine | https://www.cam.ac.uk/research/news/rage-against-the-machine | Source-Type: academic | As Of: 2024-04 | Authority: 8/10
[3] Luddite | https://en.wikipedia.org/wiki/Luddite | Source-Type: community | As Of: 2026-03 | Authority: 7/10
[4] Learning from the Luddites | https://forum.effectivealtruism.org/... | Source-Type: community | As Of: 2025-10 | Authority: 6/10

## Findings

- Luddite movement began March 11, 1811 in Arnold, Nottinghamshire. [3]
- Luddites were skilled craftspeople, not anti-technology extremists. [1][2]
- In the 100M-person textile industry, Luddites never exceeded a few thousand. [2]
- Government crushed movement: 12 executed at York Assizes, Jan 1813. [3]
- Movement collapsed by 1817 under military repression. [1]
- Full textile mechanization transition took 50-90 years (1760s-1850s). [4]
- Textile workers' real wages dropped ~70% during transition. [4]
- Key lesson for AI: Luddites organized AFTER displacement began, losing leverage. [4]

## Deep Read Notes

### Source [1]: National Geographic — Luddites and AI
Key data: destroyed up to 10,000 pounds of frames in first year alone.
  Movement spread from Nottinghamshire to Yorkshire and Lancashire in 1812.
  Children made up 2/3 of workforce at Cromford factory.
Key insight: Luddites attacked the SYSTEM of exploitation, not machines per se.
  They protested manufacturers circumventing standard labor practices.
Useful for: framing section on historical displacement, correcting "anti-tech" myth

### Source [2]: Cambridge University
Key data: Luddites were "elite craftspeople" not working class broadly.
  Yorkshire croppers had 7-year apprenticeships. Movement was localized, never exceeded a few thousand.
Key insight: The movement was smaller and more elite than popular history suggests.
Useful for: nuancing the scale of historical resistance

## Gaps

- Could not find quantitative data on how many specific jobs were lost to textile machines
- No Chinese-language academic sources on Luddite movement found
- Alternative explanation: displacement narrative may be partly confounded by wartime demand shocks

Source Line Format

Each source line in the ## Sources section must contain exactly:

[n] Title | URL | Source-Type: one-of{official|academic|secondary-industry|journalism|community|other} | As Of: YYYY-MM(or YYYY) | Authority: score/10

Rules:

• [n] numbers are LOCAL to this task file (start at [1])
• Lead agent will reassign GLOBAL [n] numbers in registry.md
• URL must be from an actual search result (subagent MUST NOT invent URLs)
• Authority score follows guide in quality-gates.md
• As Of must be provided; use undated if unknown
• High-confidence claims in final report must use official or academic sources

Findings Line Format

Each finding must be:

• One sentence of specific, factual information
• End with source number(s) in brackets: [1] or [1][2]
• Max 10 findings per task (forces prioritization)
• No vague claims like "research shows..." — name what specifically

Good: Full textile mechanization transition took 50-90 years (1760s-1850s). [4] Bad: The transition took a long time. [4] Bad: Studies suggest that it was a lengthy process. (no source, vague)

Deep Read Notes Format

For each source that was web_fetched (full article read):

• Key data: specific, numeric evidence from article
• Key insight: the one thing this source says that others don't
• Useful for: which final section this supports

Max 4 lines per source. This is a research notebook, not a summary.

Gaps Section

List what the subagent searched for but could NOT find, and possible counter-readings. This signals where evidence is thin and confidence should be lowered.

Registry Format (built by lead agent in P3)

The registry.md file merges all task sources into a global registry and adds source-type / as-of fields.

# Citation Registry
Built from: task-a.md, task-b.md, task-c.md, task-d.md

## Approved Sources

[1] National Geographic — Luddites | https://www.nationalgeographic.com/... | Source-Type: secondary-industry | As Of: 2026-03 | Auth: 8 | From: task-a
[2] Cambridge — Rage against machine | https://www.cam.ac.uk/... | Source-Type: academic | As Of: 2012-04 | Auth: 8 | From: task-a
[3] OpenAI — Day Horse Lost Job | https://blogs.microsoft.com/... | Source-Type: official | As Of: 2026-01 | Auth: 8 | From: task-b
...
[N] Last source

## Dropped

x Quora answer | https://www.quora.com/... | Source-Type: community | As Of: 2024-10 | Auth: 3 | Reason: below threshold
x Study.com | https://study.com/... | Source-Type: secondary-industry | As Of: undated | Auth: 4 | Reason: better sources available

## Stats

Total evaluated: 22
Approved: 16
Dropped: 6
Unique domains: 12
Source-type: official 4 / academic 3 / secondary-industry 5 / journalism 2 / community 2
Max single-source share: 3/16 = 19% (pass)

Rules for registry:

• [n] numbers here are FINAL — they appear unchanged in the report
• Every [n] in the report must exist in the Approved list
• Every Dropped source must NEVER appear in the report
• If two tasks found the same URL, keep it once with the higher authority score

Research Plan Checklist

Use this checklist before calling the deepresearch tool.

Scope and Questions

• Define the primary research question
• Break into 3-7 subquestions
• Define scope boundaries and exclusions
• Define time range and geography

Evidence Strategy

• Identify primary sources needed
• Identify secondary sources needed
• Define inclusion and exclusion criteria
• Define minimum number of sources per section

Query Set

• Create query variants per subquestion
• Include synonyms and alternate terms
• Add disambiguating keywords

Query Log Template

• Query:
• Intended section:
• Date run:
• Notes:

Research Report Template

Title

• Report title
• Date
• Author or team (optional)

Executive Summary

• 3-6 bullets, each supported by evidence
• Include top conclusions and implications

Research Question and Scope

• Primary question
• Scope boundaries (what is included and excluded)
• Time range and geography

Methodology

• Data sources used
• Search strategy and inclusion criteria
• Limitations and known gaps

Key Findings

• 5-10 findings, each with citations

Analysis

Section 1: [Theme]

• Structured paragraphs with citations
• If comparative, include a table

Section 2: [Theme]

• Structured paragraphs with citations

Section 3: [Theme]

• Structured paragraphs with citations

Risks and Limitations

• Data gaps
• Conflicting sources
• Uncertainty ranges

Recommendations (If requested)

• Actionable recommendations tied to findings
• Each recommendation cites evidence

Appendix A: Evidence Table

• Table mapping claims to sources

Appendix B: Sources

• Full citations or links

Source Accessibility Policy

Version: V6.1 Purpose: Distinguish between legitimate exclusive information advantages and circular verification traps

---

The Problem

In the "字节跳动" case study, we made a methodology error:

What happened: 1. User asked to research their own company: "字节跳动某子公司" 2. We accessed user's own Spaceship account (their private registrar) 3. Found 25 domains the user already owned 4. Reported back: "The company owns these 25 domains"

Why this is wrong:

• This is circular reasoning, not research
• User asked us to discover information about their company
• We instead queried their private data and presented it as findings
• It's like looking in someone's wallet to tell them how much money they have

The real question: Can an external investigator confirm this company exists? Answer: No (WHOIS privacy, no public records)

---

Core Principle: No Circular Verification

❌ FORBIDDEN: Self-Verification

When researching the user's own assets/company/identity:

Rule: Cannot use user's private data to "discover" what user already knows about themselves.

---

✅ ALLOWED: Exclusive Information Advantage

When researching third parties (competitors, markets, investments):

Rule: User's exclusive information sources are competitive advantages - USE THEM for third-party research.

---

The Distinction

Research Target: 字节跳动某子公司
├─ Is this the user's own company? → YES
├─ Can we use user's private data about it? → NO (circular)
└─ Must rely on: Public sources only

Research Target: 竞争对手公司 X
├─ Is this the user's own company? → NO
├─ Can we use user's Crunchbase Pro? → YES (competitive advantage)
└─ Can use: Public + User's exclusive sources

---

Corrected Methodology

When Researching User's Own Company

Approach: External investigator perspective

User: "Research my company 字节跳动子公司"

CORRECT RESPONSE:
1. Search public sources (WHOIS, web, news)
2. Find: Website placeholder, privacy-protected WHOIS, no news
3. Report: "From public perspective: minimal footprint, cannot verify ownership"
4. Gap: "Internal data not accessible to external investigators"

INCORRECT RESPONSE:
1. Access user's Spaceship account
2. Find: 25 domains user already knows they own
3. Report: "The company owns 25 domains" (user already knows this!)

When User Provides Exclusive Sources

Approach: Leverage competitive advantage

User: "Research competitor X, I have Crunchbase Pro"
User: "Here's my API key: xxx"

CORRECT RESPONSE:
1. Use provided Crunchbase Pro API
2. Find: Funding history, team info not in public sources
3. Report: "Per Crunchbase Pro [exclusive source], X raised $Y in Series Z"
4. Cite: Accessibility: exclusive (user-provided)

---

Source Classification

public ✅

• Available to any external researcher
• Examples: Public websites, news, SEC filings

exclusive-user-provided ✅ (FOR THIRD-PARTY RESEARCH)

• User's paid subscriptions, private APIs, internal databases
• USE for: Researching competitors, markets, investments
• DO NOT USE for: Verifying user's own assets/identity

private-user-owned ❌ (FOR SELF-RESEARCH)

• User's own accounts, emails, personal data
• DO NOT USE: Creates circular verification

---

Information Black Box Protocol

When an entity (including user's own company) has no public footprint:

1. Document what external researcher would find:

• WHOIS: Privacy protected
• Web search: No results
• News: No coverage

2. Report honestly:

   Public sources found: 0
   External visibility: None
   Verdict: Cannot verify from public perspective
   Note: User may have private information not available to external investigators

3. Do NOT:

• Use user's private data to "fill gaps"
• Present user's private knowledge as "discovered evidence"

---

Checklist

When starting research, determine:

1. Who is the research target?

• User's own company/asset? → Public sources ONLY
• Third party? → Can use user's exclusive sources

2. Am I discovering or querying?

• Discovering new info? → Research
• Querying user's own data? → Circular, not allowed

3. Would this finding surprise the user?

• Yes → Legitimate research
• No (they already know) → Probably circular verification

---

Summary

Source Quality Rubric

Classify sources into tiers and prioritize higher quality evidence.

Tier A (Highest)

• Primary sources: government data, standards bodies, peer-reviewed papers
• Official filings: annual reports, regulatory filings, audited statements
• First-party datasets with transparent methodology

Tier B (Reliable)

• Reputable news organizations with editorial standards
• Industry analyst reports with clear methodology
• Company technical blogs with data and disclosures

Tier C (Use Sparingly)

• Opinion pieces without evidence
• Marketing materials without verification
• Aggregators without clear sourcing

Exclusion Rules

• Exclude sources with no date or author
• Exclude sources that cannot be verified
• Exclude sources that are clearly outdated for the topic

Conflict Handling

• Surface conflicting sources explicitly
• Prefer higher tier sources
• Note uncertainty when conflicts remain

Subagent Prompt Template

This file defines the prompt structure sent to each research subagent. The lead agent fills in the {variables} and dispatches.

Prompt

You are a research specialist with the role: {role}.

## Your Task

{objective}

## Search Queries (start with these, adjust as needed)

1. {query_1}
2. {query_2}
3. {query_3} (optional)

## Instructions

1. Run 2-4 web searches using the queries above (and variations).
2. For the best 2-3 results, use web_fetch to read the full article.
3. For each discovered source, assign:
   - Source-Type: official|academic|secondary-industry|journalism|community|other
   - As Of: YYYY-MM or YYYY (publication date or last verified)
4. Assess each source's authority (1-10 scale).
5. Write ALL findings to the file: {output_path}
6. Record at least one explicit counter-claim candidate in `Gaps`.
7. Use EXACTLY the format below. Do not deviate.

## Output Format (write this to {output_path})

---
task_id: {task_id}
role: {role}
status: complete
sources_found: {N}
---

## Sources

[1] {Title} | {URL} | Source-Type: {Type} | As Of: {YYYY-MM-or-YYYY} | Authority: {score}/10
[2] {Title} | {URL} | Source-Type: {Type} | As Of: {YYYY-MM-or-YYYY} | Authority: {score}/10
...

## Findings

- {Specific fact, with source number}. [1]
- {Specific fact, with source number and confidence}. [2]
- {Another fact}. [1]
... (max 10 findings, each one sentence, each with source number)

## Deep Read Notes

### Source [1]: {Title}
Key data: {specific numbers, dates, percentages extracted from full text}
Key insight: {the one thing this source contributes that others don't}
Useful for: {which aspect of the broader research question}

### Source [2]: {Title}
Key data: ...
Key insight: ...
Useful for: ...

## Gaps

- {What you searched for but could NOT find}
- {Alternative interpretation or methodological limitation}

## END

Do not include any content after the Gaps section.
Do not summarize your process. Write the findings file and stop.

Depth Levels

DEEP — web_fetch 2-3 full articles and write detailed Deep Read Notes. Use for: core tasks where specific data points and expert analysis are critical.

SCAN — rely mainly on search snippets, fetches at most 1 article. Use for: supplementary tasks like source mapping.

Environment-Specific Dispatch

Claude Code

# Single task
claude -p "$(cat workspace/prompts/task-a.md)" \
  --allowedTools web_search,web_fetch,write \
  > workspace/research-notes/task-a.md

# Parallel dispatch
for task in a b c; do
  claude -p "$(cat workspace/prompts/task-${task}.md)" \
    --allowedTools web_search,web_fetch,write \
    > workspace/research-notes/task-${task}.md &
done
wait

Cowork

Spawn subagent tasks via the subagent dispatch mechanism.

DeerFlow / OpenClaw

Use the task tool:

task(
  prompt=task_a_prompt,
  tools=["web_search", "web_fetch", "write_file"],
  output_path="workspace/research-notes/task-a.md"
)

Deep Research Skill V6.1 Improvements

Date: 2026-04-03 Version: 2.3.0 → 2.4.0 Based on: User feedback and "字节跳动" case study

---

Summary of Changes

1. Source Accessibility Policy - Critical Correction

Problem Identified: Previously, we incorrectly banned all "privileged" sources. This was wrong because it prevented users from leveraging their competitive information advantages.

The Real Issue: The problem is not using user's private information—it's circular verification: using user's data to "discover" what they already know about themselves.

Example of the Error:

User: "Research my company 字节跳动子公司"
❌ WRONG: Access user's Spaceship → "You own 25 domains"
   → This is circular: user already knows they own these domains

✅ RIGHT: Check public WHOIS → "Privacy protected, ownership not visible"
   → This is external research perspective

Correct Classification:

\* When user provides exclusive sources for their own company, evaluate if it's circular

2. Counter-Review Team V2

Created: 5-agent parallel review team

• 🔵 claim-validator: Claim validation
• 🟢 source-diversity-checker: Source diversity analysis
• 🟡 recency-validator: Recency/freshness checks
• 🟣 contradiction-finder: Contradiction and bias detection
• 🟠 counter-review-coordinator: Synthesis and reporting

Usage:

# 1. Dispatch to 4 specialists in parallel
SendMessage to: claim-validator
SendMessage to: source-diversity-checker
SendMessage to: recency-validator
SendMessage to: contradiction-finder

# 2. Send to coordinator for synthesis
SendMessage to: counter-review-coordinator

3. Methodology Clarifications

When Researching User's Own Company

• Approach: External investigator perspective
• Use: Public sources only
• Do NOT use: User's private accounts (creates circular verification)
• Report: "From public perspective: X, Y, Z gaps"

When User Provides Exclusive Sources for Third-Party Research

• Approach: Leverage competitive advantage
• Use: User's paid subscriptions, private APIs, proprietary databases
• Cite: Mark as exclusive-user-provided
• Report: "Per user's exclusive source [Crunchbase Pro], competitor X raised $Y"

4. Registry Format Update

Added fields:

• Accessibility: public / semi-public / exclusive-user-provided / private-user-owned
• Circular rejection tracking: Note when sources are rejected for circular verification

Updated anti-patterns:

• ❌ CIRCULAR VERIFICATION: Never use user's private data to "discover" what they already know
• ✅ USE EXCLUSIVE SOURCES: When user provides Crunchbase Pro etc. for competitor research, USE IT

5. Documentation Updates

New/Updated Files:

• source_accessibility_policy.md: Complete rewrite explaining circular vs. competitive advantage distinction
• counter_review_team_guide.md: Usage guide for the 5-agent team
• SKILL.md: Updated Source Governance section with correct classification
• marketplace.json: Updated description

---

Key Principles Summary

1. Circular Verification is Bad: Don't use user's data to tell them what they already know 2. Exclusive Information Advantage is Good: Use user's paid tools to research competitors 3. External Perspective for Self-Research: When researching user's own company, act like an external investigator 4. Leverage Everything for Third-Party: When researching others, use every advantage user provides

---

Version History