Deep Research Framework
A structured deep-research skill for producing rigorous, balanced, and actionable research reports on any topic.
Overview
This skill enforces a disciplined research methodology. It is NOT a simple "search and summarize" tool.
It demands evidence quality control, counterevidence inclusion, and honest flagging of gaps.
Outputs:
- Structured Research Report — full-depth analysis organized by dimensions
- One-Page Summary Card — key findings, confidence levels, and action items
Phase 0 — Pre-Research Clarification
Before starting research, confirm these inputs from the user or infer them from context:
| Input | Required | Default if absent |
|---|
| Research topic | Yes | — |
| Target audience | Recommended | "decision maker / general professional" |
| Key dimensions to focus on | Optional | Use standard dimension set (see references) |
| Time scope | Optional | "recent 1–3 years unless specified" |
| Depth level | Optional | "standard" (balanced breadth and depth) |
If the topic is ambiguous, ask ONE clarifying question before proceeding.
Phase 1 — Research Design
1.1 Define the Research Objective
State in 1–2 sentences: What question is this research answering, and what decision or understanding will it enable?
1.2 Define the Audience
State who will read this: their background, what they already know, and what they need to walk away with.
1.3 Scope and Constraints
- Time scope: Which period does the research cover?
- Geographic scope: Global? China? Specific region?
- Exclusions: What is explicitly out of scope?
1.4 Decompose into Research Dimensions
Break the topic into 4–7 dimensions. For each dimension:
- Write 2–4 key questions that need answering
- Assign a priority: Core / Supporting / Context
Use references/dimension-templates.md for standard dimension sets by research type (industry, technology, company, product, market).
Phase 2 — Multi-Source Research Execution
2.1 Source Strategy
For each dimension, gather from at least 2–3 distinct source types:
| Source Type | Examples | Default Credibility |
|---|
| Primary data | Official reports, regulatory filings, academic papers | ★★★★★ |
| Industry reports | McKinsey, Gartner, IDC, CB Insights, Frost & Sullivan | ★★★★ |
| News & media | Reuters, Bloomberg, FT, 36kr, 财新, 虎嗅 | ★★★ |
| Expert commentary | Interviews, conference talks, analyst notes | ★★★ |
| Community & forums | Reddit, Zhihu, GitHub Issues, industry forums | ★★ |
| AI-generated synthesis | LLM summaries without cited sources | ★ |
See references/credibility-rubric.md for detailed credibility rating rules.
2.2 Search Execution
For each research dimension:
- Formulate targeted search queries (try multiple angles)
- Retrieve results using web_search or available tools
- For each piece of evidence, record: claim, source, credibility rating, date
2.3 Counterevidence Requirement (MANDATORY)
For every major claim in the report, actively search for contradicting evidence:
- If counterevidence exists: include it alongside the main claim, labeled
[反面证据]
- If no counterevidence found: note
[未发现明显反面证据，可信度待验]
- NEVER omit counterevidence to make the report cleaner or more conclusive
2.4 Conflicting Information Handling
When two credible sources contradict each other:
- Present BOTH positions side-by-side
- Note the source, date, and credibility of each
- Do NOT pick a side unless evidence weight is clearly asymmetric
- Label the conflict:
[矛盾信息 — 并列呈现]
2.5 Information Gap Flagging (MANDATORY)
When a key question cannot be answered from available sources:
- Mark it clearly:
[信息不足 — 无法得出可靠结论]
- State what type of source would be needed to fill the gap
- NEVER fabricate data, statistics, or citations to fill gaps
Phase 3 — Synthesis and Analysis
3.1 Evidence Weighting
Synthesize findings dimension by dimension:
- Higher-credibility sources take precedence
- Recent data (within 12 months) takes precedence over older data
- Consensus across multiple independent sources strengthens confidence
3.2 Insight Generation
Go beyond summarizing to identify:
- Key patterns and trends
- Surprising or counterintuitive findings
- Implications for the research objective
- Open questions that remain after research
3.3 Confidence Levels
Assign a confidence level to each major finding:
| Level | Meaning |
|---|
| 高 (High) | Multiple independent ★★★★+ sources agree |
| 中 (Medium) | Some evidence, limited cross-validation |
| 低 (Low) | Single source, low credibility, or conflicting signals |
| 未知 (Unknown) | Insufficient data, explicitly flagged |
Phase 4 — Output Delivery
4.1 Full Research Report
Follow the template in assets/report-template.md. Structure:
- 研究概览 — objective, audience, scope, date
- 执行摘要 — 5–7 bullet key findings with confidence levels
- 维度分析 — one section per dimension with:
- Key findings
- Supporting evidence (with source + credibility)
- Counterevidence
- Confidence level
- Information gaps (if any)
- 矛盾与争议 — consolidated list of conflicting information
- 综合洞察 — cross-dimension synthesis and implications
- 信息缺口清单 — all flagged gaps in one place
- 参考资料 — all sources with credibility ratings
4.2 One-Page Summary Card
Follow the template in assets/summary-card-template.md. Deliver after the full report.
The card contains:
- Topic + audience + date (header)
- 3 most important findings (with confidence)
- 1 most important counterpoint
- 1 key uncertainty / gap
- Suggested next actions (if applicable)
- Overall research confidence score (High / Medium / Low)
Quality Checklist (self-check before delivery)
Before presenting the output, verify:
Important Rules
- No fabrication: If data doesn't exist, say so. Never invent statistics, quotes, or citations.
- No false certainty: Do not present uncertain findings as definitive.
- Counterevidence is mandatory: A report without counterevidence is incomplete.
- Conflicting sources must coexist: Do not silently pick one side.
- Credibility must be visible: Every claim should be traceable to a rated source.
