| name | deep-research |
| description | Comprehensive research framework that combines web search, content analysis, source verification, and iterative investigation to conduct in-depth research on any topic. Use when you need to perform thorough research with multiple sources, cross-validation, and structured findings. |
Deep Research Framework
Overview
The Deep Research skill provides a systematic approach to conducting thorough investigations on any topic. It combines multiple tools and methodologies to gather, analyze, verify, and synthesize information.
Core Components
1. Research Planning
- Define research objectives
- Identify key questions
- Establish search criteria
- Determine validation requirements
2. Information Gathering
- Multi-source web search
- Content extraction from various formats
- Source diversity verification
- Temporal relevance assessment
3. Analysis & Synthesis
- Cross-reference multiple sources
- Identify patterns and contradictions
- Evaluate source credibility
- Organize findings systematically
4. Validation & Verification
- Fact-checking against authoritative sources
- Cross-validation of claims
- Identify potential biases
- Assess information reliability
Research Workflow
Phase 1: Initial Investigation
0. Perspective Alignment (视角模式, optional)
⚠️ This step only executes when a perspective/analytical framework is specified. Skip entirely when no perspective is given.
When the user specifies a perspective (e.g., "用 Serenity 的框架", "chokepoint 视角", "用某投资人的视角"):
-
Load perspective framework:
- If referencing an installed persona Skill (e.g.,
serenity-reply): read its SKILL.md, extract mental models and decision heuristics
- If referencing a built-in framework: use the pre-defined core questions
- If custom: distill 3-5 core questions from user description
-
Translate framework into research dimensions: For each mental model, define:
- Core question it asks
- Specific search terms it suggests
- What findings it would prioritize
-
Perspective-driven search strategy: Augment broad search with perspective-specific queries:
- Example (chokepoint perspective): add "monopoly chokepoint", "sole supplier", "irreplaceable", "barrier to entry"
- Example (NVIDIA signal perspective): add "NVIDIA investment", "NVIDIA roadmap", "NVIDIA supply chain"
- Example (info asymmetry perspective): add "institutional buying", "small cap undervalued", "overlooked"
-
Output: Internal perspective config (not shown to user):
Perspective: [name]
Core questions: [2-5 questions]
Search dimensions: [custom search terms per dimension]
Priority rules: [which findings rank higher]
Confidence labels: [4-tier system definitions]
-
Topic Analysis
- Clarify research scope
- Identify key concepts and terms
- Define specific questions to answer
- If perspective mode: align questions with perspective's core questions
-
Broad Search
- Use
web_search to identify major sources
- Gather diverse perspectives
- Map the landscape of available information
- If perspective mode: supplement broad search with perspective-specific search terms (see Step 0)
-
Source Prioritization
- Rank sources by authority and relevance
- Identify primary vs. secondary sources
- Note publication dates and context
Phase 2: Deep Dive
-
Detailed Content Extraction
- Use
web_fetch to retrieve full articles/pages
- Extract key information systematically
- Maintain source attribution
-
Cross-Reference Analysis
- Compare claims across multiple sources
- Identify agreements and disagreements
- Note inconsistencies for further investigation
-
Expert Sources
- Seek academic papers, expert opinions
- Look for peer-reviewed sources
- Identify recognized authorities on the topic
Phase 3: Synthesis & Validation
-
Pattern Recognition
- Identify consistent themes across sources
- Highlight areas of disagreement
- Note gaps in available information
-
Fact Verification
- Cross-check claims against authoritative sources
- Verify dates, statistics, and attributions
- Identify potential misinformation
-
Bias Assessment
- Evaluate source objectivity
- Identify potential conflicts of interest
- Consider temporal context of information
Phase 4: Report Generation — ⚠️ MANDATORY POLISHER + VISUALIZATION STEP
⚠️ CRITICAL — Two mandatory steps before final output:
- report-polisher-zh 研报润色 — MUST run before Mermaid conversion
- Mermaid 可视化 — ASCII diagrams MUST be converted to Mermaid
Both steps are mandatory. Do not skip.
Step 1: report-polisher-zh 研报润色(必须在 Mermaid 之前执行)
在生成 Mermaid 图表之前,先对报告全文进行专业研报润色与 AI 痕迹清除。具体流程:
- 加载 report-polisher-zh skill:读取
~/.qclaw/workspace/skills/report-polisher-zh/SKILL.md
- 研报结构检查(按清单逐项核查):
- 标题规范性(含公司/行业名称、核心观点或投资评级)
- 摘要完整性(核心逻辑、关键数据、投资建议)
- 章节结构(行业概况、公司分析、财务分析、风险提示等)
- 数据引用规范(来源标注、时间节点、数值单位统一)
- 图表标注规范(编号连续、标题清晰、来源标注)
- 风险提示完整性(主要风险、具体描述、风险等级)
- 朱雀检测对抗三步法(针对统计概率分布,非口语化):
- 术语注入:用低频专业术语替代高频通俗词("很多"→"众多"、"但是"→"然而"),使用行业特有表达("赛道"、"估值锚"、"业绩兑现")
- 信息压缩:提高信息密度,删除冗余修饰词,合并同类信息,每句话必须承载有效信息
- 句式打散:长短句极端混合,打破节奏规律性,避免"首先...其次...最后..."套路结构
- 改写强度选择(根据 AI 痕迹程度):
- 轻度(10-20%):术语替换 + 句式微调
- 中度(30-50%):段落重组 + 信息压缩
- 深度(60-80%):结构重构 + 逻辑重排
- 保留 Mermaid 代码块:润色过程中不修改 Mermaid 图表代码块,只处理文字描述部分
⚠️ 无效方法(已验证失败,禁止使用):
- ❌ 口语化改写
- ❌ 情感注入
- ❌ 简单同义词替换
- ❌ 增加"我认为"等主观表达
有效策略:
- ✅ 术语注入
- ✅ 信息压缩
- ✅ 句式打散
- ✅ 适度瑕疵注入
- ✅ 个人表达习惯注入
⚠️ 注意:研报润色和 Mermaid 优化是两个独立步骤。润色处理文字内容;Mermaid 优化处理图表。二者顺序不能颠倒。
When the research involves any of the following, you MUST invoke the mermaid-diagrams skill and convert ASCII/text diagrams into proper Mermaid syntax:
- 📊 Architecture diagrams (公司架构、业务结构、股权结构)
- 🔄 Process flows (业务流程、重整理赔流程、重组时间线)
- 🏢 Organizational charts (股权结构图、集团子公司关系图)
- 📈 Market comparison tables (竞争格局矩阵)
- 🔗 Relationship diagrams (供应链关系、上下游关系、概念股图谱)
- 📋 Timeline diagrams (发展历程、重大事件时间轴)
Conversion rules (ASCII → Mermaid):
| ASCII 类型 | Mermaid 类型 | 理由 |
|---|
| 树状/层级文本图 | flowchart TD/LR | 展示层级关系和决策路径 |
| 股权结构文字 | flowchart TB + 子图 subgraph | 清晰展示控制链 |
| 业务流程步骤 | flowchart LR + 决策节点 | 展示流程与分支 |
| 公司沿革时间线 | flowchart LR 或 gantt | 展示时间序列 |
| 竞争格局对比 | quadrantChart 或表格 | 不适合 Mermaid → 保留 Markdown 表格 |
| 架构分层图 | C4 diagram 或 flowchart | 系统/业务架构 |
⚠️ 关键约束(Mermaid 图表规则):
- 节点数 ≤ 15 个(超出则拆分为多个图)
- 每条连线必须有文字标签
- 必须有标题/图注
- 禁止用 Mermaid 画数据图表(柱状图/折线图等)
- 产出为
.mmd 文件或嵌入 Markdown 的 Mermaid 代码块
🎨 公众号风格配置(柔和马卡龙版)—— 强制应用
所有 Mermaid 图表必须使用以下配置,确保颜色温柔、排版舒展,完全适配公众号阅读体验:
%%{init: {
'theme': 'base',
'themeVariables': {
'primaryColor': '#FFE5D9',
'primaryTextColor': '#5D4E37',
'primaryBorderColor': '#E8C4B8',
'lineColor': '#B8A99A',
'secondaryColor': '#D4E7D4',
'secondaryTextColor': '#4A6741',
'secondaryBorderColor': '#A8C8A8',
'tertiaryColor': '#E8DFF5',
'tertiaryTextColor': '#5D4E6D',
'tertiaryBorderColor': '#C8B8D8',
'background': '#FFFBF5',
'mainBkg': '#FFF5EE',
'nodeBorder': '#E8D5C4',
'clusterBkg': '#F5F0EB',
'clusterBorder': '#D4C4B8',
'titleColor': '#6B5B4F',
'edgeLabelBackground': '#FFF9F0',
'fontFamily': 'PingFang SC, Microsoft YaHei, sans-serif',
'fontSize': '15px'
},
'flowchart': {
'curve': 'basis',
'padding': 20,
'nodeSpacing': 50,
'rankSpacing': 80,
'diagramPadding': 30
}
}}%%
马卡龙色板(节点专用):
| 色块 | 色值 | 适用场景 |
|---|
| 🍑 蜜桃粉 | #FFE5D9 | 核心节点、重要结论 |
| 🌿 抹茶绿 | #D4E7D4 | 产业链上游、供给端 |
| 💜 薰衣紫 | #E8DFF5 | 科技创新、AI相关 |
| 🍋 柠檬黄 | #FFF3CD | 数据指标、业绩表现 |
| 🌊 天空蓝 | #D6EAF8 | 需求端、下游应用 |
| 🌸 樱花粉 | #FADBD8 | 风险提示、负面因素 |
公众号适配排版规范:
- 节点文字:中文优先,单节点≤20字,分行显示
- 连线标签:必须添加,说明关系(如"驱动"、"传导"、"受益")
- 子图分组:使用
subgraph 划分逻辑区域,背景色用 #F8F4F0
- 整体布局:优先
TD(上下)或 LR(左右),避免斜线交叉
- 节点形状:
- 矩形
[] → 常规节点
- 圆角
() → 起点/终点
- 菱形
{} → 判断/决策
- 圆柱
[()] → 数据库/存储
Step-by-step conversion workflow:
Step 1: 识别报告中的 ASCII 图/文字架构图
↓
Step 2: 判断最适合的 Mermaid 类型 (flowchart / C4 / gantt / etc.)
↓
Step 3: 读取 mermaid-diagrams skill 获取该类型的标准语法
↓
Step 4: 逐节点转换(节点 label 保持中文,≤ 40字符)
↓
Step 5: 验证 Mermaid 语法(无断行语法错误)
↓
Step 6: 嵌入 Markdown 报告(或输出独立 .mmd 文件)
📋 公众号风格 Mermaid 示例(完整模板):
%%{init: {
'theme': 'base',
'themeVariables': {
'primaryColor': '#FFE5D9',
'primaryTextColor': '#5D4E37',
'primaryBorderColor': '#E8C4B8',
'lineColor': '#B8A99A',
'secondaryColor': '#D4E7D4',
'secondaryTextColor': '#4A6741',
'secondaryBorderColor': '#A8C8A8',
'tertiaryColor': '#E8DFF5',
'tertiaryTextColor': '#5D4E6D',
'tertiaryBorderColor': '#C8B8D8',
'background': '#FFFBF5',
'mainBkg': '#FFF5EE',
'nodeBorder': '#E8D5C4',
'clusterBkg': '#F5F0EB',
'clusterBorder': '#D4C4B8',
'titleColor': '#6B5B4F',
'edgeLabelBackground': '#FFF9F0',
'fontFamily': 'PingFang SC, Microsoft YaHei, sans-serif',
'fontSize': '15px'
},
'flowchart': {
'curve': 'basis',
'padding': 20,
'nodeSpacing': 50,
'rankSpacing': 80,
'diagramPadding': 30
}
}}%%
flowchart TD
subgraph 上游供给["🏭 上游供给端"]
A["原材料\n涨价压力"] --> B["产能出清\n供给收缩"]
end
subgraph 中游传导["⚙️ 中游制造端"]
B --> C["成本传导\n提价落地"]
C --> D["毛利修复\n业绩改善"]
end
subgraph 下游需求["🛒 下游需求端"]
E["AI算力爆发\n需求激增"] --> C
F["地缘冲突\n避险需求"] --> G["资源品\n价格飙升"]
end
G --> A
subgraph A股机会["📈 A股投资机会"]
D --> H["涨价链龙头\n业绩兑现"]
G --> I["资源股\n通胀受益"]
end
style A fill:#FFE5D9,stroke:#E8C4B8,color:#5D4E37
style E fill:#E8DFF5,stroke:#C8B8D8,color:#5D4E6D
style F fill:#D6EAF8,stroke:#A8C8D8,color:#4A678D
style H fill:#D4E7D4,stroke:#A8C8A8,color:#4A6741
style I fill:#FFF3CD,stroke:#E8D8A8,color:#6B5B4F
⚠️ 每个图表开头必须包含完整的 %%{init: {...}}%% 配置块,否则无法应用马卡龙风格。
Phase 4 完整输出清单:
- ✅ Humanizer-zh 去AI化(文字内容已去AI味,保留 Mermaid 代码块)
- ✅ Executive Summary(执行摘要,2-3句,人性化表达)
- ✅ Mermaid 可视化图表(所有 ASCII 架构图已转换)
- ✅ Structured findings(按主题组织,去AI化)
- ✅ Source evaluation(来源可信度评估)
- ✅ Limitations(研究局限性,去AI化)
- ✅ Remaining questions(待深入研究的问题)
1. Structured Summary
- Executive summary of key findings
- Detailed findings organized by theme
- Supporting evidence for each claim
2. Mermaid Visualization (MANDATORY)
- Convert ALL ASCII/text architecture diagrams to Mermaid
- Include at minimum: corporate structure, business overview, key timelines
- Use appropriate diagram type per content (see conversion table above)
3. Source Evaluation
- Assessment of source credibility
- Identification of limitations
- Confidence levels for different claims
4. Remaining Questions
- Areas requiring further investigation
- Conflicting information needing resolution
- Gaps in current knowledge
Phase 5: Mermaid → Instagram 信息图导出(报告完成后执行)
触发条件:报告已通过 Phase 4 完成润色 + Mermaid 嵌入,Markdown 文件已保存到磁盘后执行。
核心原则:只生成图片,严禁修改报告 Markdown 原文(不使用 --update 参数)。
Step 1: 定位报告文件与输出目录
- 确认最终报告的
.md 文件路径(Phase 4 保存位置)
- 图片输出目录 = 报告文件所在目录(同级目录,不另建子文件夹)
- 例如:报告路径为
~/Desktop/A股深度产研报告/XX报告-20260412.md
- 则图片输出到
~/Desktop/A股深度产研报告/
- 通过
-o <报告所在目录> 参数指定
Step 2: 执行 Mermaid 批量转图
cd ~/projects/flowchart-to-instagram
python scripts/md2images.py "<报告.md完整路径>" -o "<报告所在目录>" --prefix fig_
参数说明:
| 参数 | 作用 | 注意 |
|---|
<报告.md完整路径> | 指定要处理的 Markdown 文件 | 绝对路径或相对路径均可 |
-o "<报告所在目录>" | 图片输出到报告同级目录 | 与报告同目录,方便查找 |
--prefix fig_ | 图片文件名前缀(如 fig_1.png、fig_2.png) | 可自定义,默认为 chart_ |
⚠️ 禁止事项:
- ❌ 禁止使用
--update 参数(会将 Mermaid 代码块替换为图片链接,破坏报告原文)
- ❌ 禁止修改报告 Markdown 文件(图片是报告的附属产物,报告原文保持不变)
预期输出:
- 报告同级目录下生成
fig_1.png、fig_2.png、... 等图片文件
- 图片数量 = 报告中 Mermaid 代码块的数量
- 图片内容为 Instagram 风格信息图,适合公众号投放
Step 3: 验证输出
- 检查输出目录下是否生成了预期数量的 PNG 文件
- 确认报告 Markdown 原文未被修改(Mermaid 代码块完整保留)
- 向用户报告:图片数量、文件名、存放路径
报告最终交付物清单(Phase 4 + Phase 5)
| # | 交付物 | 格式 | 路径 |
|---|
| 1 | 完整研究报告(含 Mermaid 代码块) | .md | 报告所在目录 |
| 2 | Mermaid 信息图(Instagram 风格) | .png | 与报告同目录(fig_N.png) |
Tools Integration
Web Research
web_search: Initial broad search to identify sources
web_fetch: Retrieve detailed content from specific URLs
browser: For complex sites or when web_fetch fails
Content Processing
read: Process downloaded content or documents
write: Create structured research notes
edit: Refine and organize findings
Mermaid Visualization (MANDATORY — Phase 4)
⚠️ Use mermaid-diagrams skill for ALL diagram generation.
ASCII diagrams in reports are a quality failure. Every architecture,
process, or relationship diagram must be rendered as proper Mermaid.
mermaid-diagrams skill: Read the skill file to get standard syntax for
flowchart, C4, gantt, sequenceDiagram, classDiagram, erDiagram
- Output:
.mmd files or mermaid code blocks embedded in Markdown
- See Phase 4 rules for node count limit (≤ 15), labeling rules, and type selection
🎨 公众号风格(强制应用):
- 所有图表必须使用 马卡龙色系(蜜桃粉
#FFE5D9、抹茶绿 #D4E7D4、薰衣紫 #E8DFF5 等)
- 配置块
%%{init: {...}}%% 必须置于图表最开头
- 字体:
PingFang SC, Microsoft YaHei, sans-serif,字号 15px
- 布局舒展:
nodeSpacing: 50, rankSpacing: 80, padding: 20
- 节点分行显示,连线必须有中文标签
Memory & Organization
memory_get / memory_search: Reference previous research
write: Create persistent research records
- Structured file organization for findings
Research Quality Standards
Source Diversity
- Include multiple perspectives on controversial topics
- Balance popular and academic sources
- Include international viewpoints when relevant
- Seek primary sources when possible
Temporal Relevance
- Prioritize recent information for fast-changing topics
- Consider historical context for trend analysis
- Note when information was published
- Flag potentially outdated information
Authority Assessment
- Prioritize peer-reviewed academic sources
- Consider author credentials and institutional affiliation
- Check for potential conflicts of interest
- Verify organizational reputation
Iterative Research Approach
Cycle 1: General Overview
- Broad search to understand the topic landscape
- Identify key terms, concepts, and stakeholders
- Establish initial research questions
Cycle 2: Focused Investigation
- Targeted searches based on initial findings
- Deep dive into specific aspects
- Begin synthesis of information
Cycle 3: Validation & Refinement
- Verify key claims across multiple sources
- Resolve contradictions
- Refine understanding based on evidence
Cycle 4: Synthesis & Reporting
- Combine findings into coherent narrative
- 🔴 Humanizer-zh 去AI化(MUST run before Mermaid, text content only)
- 🔴 Convert ALL ASCII diagrams to Mermaid (mermaid-diagrams skill is MANDATORY)
- Identify remaining uncertainties
- Prepare final research report
- Quality gate: (1) Has text been humanized? (2) Does the report contain any ASCII art/box diagrams that should be Mermaid?
Cycle 5: Mermaid 信息图导出
- 报告保存到磁盘后,使用
flowchart-to-instagram skill 将 Mermaid 代码块转为 PNG
- 图片输出到报告同目录,不修改报告原文(见 Phase 5 规范)
- 验证 PNG 生成数量与 Mermaid 代码块数量一致
Output Structure
Research Report Template
# [Research Topic] - Deep Research Report
> **研报润色说明**:本报告已通过 report-polisher-zh 润色,文字表达符合专业研报规范,
> Mermaid 图表代码块在润色过程中已保留原样。
## Executive Summary
[2-3 sentence summary of key findings — humanized, natural tone]
## Research Questions
[Specific questions investigated]
## Methodology
[Description of research approach and tools used]
## Key Findings
[Main discoveries organized by theme — humanized language]
## [REQUIRED] Mermaid Visualizations
[Convert all ASCII diagrams to Mermaid code blocks.
At minimum include: corporate structure, business overview, key event timeline.
Mermaid code blocks go in this section. See Phase 4 Mermaid rules.]
## Supporting Evidence
[Evidence supporting each finding with sources]
## Contradictions/Debates
[Areas of disagreement among sources]
## Source Credibility Assessment
[Evaluation of information sources]
## Limitations
[Identified limitations in research — humanized language]
## Perspective Review (视角审查, perspective mode only)
> ⚠️ This section only appears when a perspective/analytical framework was specified.
### Framework-by-Framework Review
[For each mental model in the perspective framework:]
- What findings does this model explain?
- What questions does this model raise that the report didn't cover?
- What are the limitations of applying this model here?
### Conclusion Confidence Table
| Conclusion | Confidence Tier | Notes |
|-----------|----------------|-------|
| [conclusion] | Direct Evidence / Pattern Induction / Framework Extrapolation / Insufficient Evidence | [source/logic] |
> ⚠️ Investment disclaimer: Framework perspective, not investment advice. DYOR.
## Further Research Needed
[Questions requiring additional investigation]
📦 最终交付物(Phase 5 补充):
报告 Markdown 保存到磁盘后,自动执行 Phase 5,将报告中的所有 Mermaid 代码块导出为 PNG 图片:
- 图片格式:
fig_1.png、fig_2.png、...
- 图片路径:与报告
.md 文件同目录
- 报告 Markdown 原文不被修改(Mermaid 代码块保留在报告中)
⚠️ REMINDER: Before finalizing the report:
- ALWAYS invoke
report-polisher-zh skill and polish text content first (preserve Mermaid code blocks)
- THEN invoke
mermaid-diagrams skill and convert any ASCII/text architecture diagrams to proper Mermaid syntax.
- FINALLY invoke Phase 5 — use
flowchart-to-instagram skill to export Mermaid diagrams as PNG images to the same directory as the report, without modifying the report markdown.
See Phase 4 → report-polisher-zh + Mermaid Visualization sections for full rules.
See Phase 5 → Mermaid → Instagram 信息图导出 for image generation rules.
Use Cases
Academic Research
- Literature reviews
- Topic exploration
- Source verification
Business Intelligence
- Market analysis
- Competitive research
- Technology trends
Fact Checking
- Claim verification
- Misinformation identification
- Source credibility assessment
Personal Learning
- Deep topic exploration
- Concept clarification
- Question resolution
Quality Assurance
- Always verify critical claims against multiple sources
- Flag information that seems unreliable
- Maintain skepticism toward sensational claims
- Prioritize authoritative sources over anonymous ones
- Document all sources for verification purposes