research-survey（研究调查）

Research Survey

Generates high-quality, survey-grade literature reviews from papers collected by paper-navigator.

paper-navigator (collect 30-120 papers)
    ↓
Stage 1: Generate Outline (query-type adaptive structure)
    ↓
Stage 2: Draft Survey (outline + top-30 papers)
    ↓
Stage 3: Expand Sections (draft + all papers, section-by-section)
    ↓
Stage 4: Generate Section Summaries
    ↓
Stage 5: Refine Summary Sections (Abstract/Intro/Conclusion)
    ↓
Stage 6: Assemble + References

When to Use

• User asks for a "literature review", "survey", "field overview", or "systematic review"
• User has collected papers and wants them synthesized into a structured report
• User wants to understand the full landscape of a research field

When NOT to Use

• Finding papers → use paper-navigator first, then come here
• Generating research ideas → use research-ideation
• Writing a Related Work section for a paper → use paper-writing

Dependency: paper-navigator

This skill requires papers as input. If the user hasn't provided papers, first invoke paper-navigator (Workflow 1, target 30-120 papers) to collect them.

CRITICAL: All paper discovery MUST use the paper-navigator skill and its scripts (scholar_search, citation_traverse, arxiv_monitor, recommend, etc.). Using WebSearch, WebFetch, or any generic web search tool for finding papers is PROHIBITED. Generic web search cannot access Semantic Scholar, citation graphs, or academic recommendation systems. Only paper-navigator provides the academic search infrastructure needed for survey-quality literature collection.

Stage 1: Generate Outline

This is a two-phase process. Different fields have different survey conventions — a clinical systematic review looks nothing like a CS methods survey. First generate a domain-appropriate template, then create the detailed outline.

Phase 1A: Generate Domain-Specific Survey Template

Before outlining, identify the field and adapt the structure:

1. Identify the field from the user's goal and collected papers
2. Select section names and organization logic using the field-specific conventions in assets/survey-template.md (e.g., medicine organizes by intervention type and follows PRISMA; chemistry organizes by reaction class; social sciences organize by theoretical perspective)
3. Add field-specific sections (e.g., Risk of Bias Assessment for medicine, Structure-Property Relationships for materials, Ethical Considerations for human-subjects research)
4. Determine comparison table dimensions appropriate to the field

Phase 1B: Create Detailed Outline

With the domain-specific template as the framework, generate the outline:

Query Type Classification

Outline Requirements

The outline is NOT a simple heading list — it's a blueprint with meta-instructions for each section. For each ## Section:

• Include [Instruction:...] specifying what the section must contain
• Specify required tables with field-appropriate columns
• For main body sections: mandate taxonomy by underlying principle/mechanism, NOT chronology
• Include any field-required elements (e.g., PRISMA flowchart for medical systematic reviews, mathematical formalism for physics)

See references/survey-methodology.md for full outline generation rules and assets/survey-template.md for field-specific conventions.

Stage 2: Draft Survey

Generate a complete draft from the outline using the top-30 most relevant papers.

• Use numbered citations [1], [2, 3] throughout
• Follow the outline's meta-instructions strictly
• Each methods section must build a taxonomy and include comparison tables
• Problem definition must include LaTeX formalization ($$...$$)

Stage 3: Expand Sections

Expand each non-summary section using all collected papers (30-120). This is where survey-grade depth is achieved.

Section Expansion Targets

Expansion Rules

1. Thematic coherence: Keep same themes and narrative flow as draft — don't introduce unrelated topics
2. Cite comprehensively: Use as many relevant papers from the full collection as possible
3. Survey-grade depth: Multi-paragraph technical narratives per method family, not shallow bullet points
4. For each paradigm/method family, include:

- Technical narrative: How it works, theoretical assumptions, nuances between papers - Critical analysis: Why effective, trade-offs, failure modes - Comparative analysis table: Method | Core Mechanism | Key Advantage | Limitation | Performance

Stage 4: Generate Section Summaries

After all content sections are expanded, generate a condensed summary for each major section:

1. Summarize each expanded section in 150-300 words
2. Preserve the key taxonomy, representative methods, and main trade-offs
3. Keep citation anchors so later summary sections remain grounded

These section summaries become the shared context for the final abstract, introduction, and conclusion.

Stage 5: Refine Summary Sections

After all content sections are expanded, refine the summary sections (Abstract, Introduction, Conclusion):

1. Use all section summaries as context to rewrite Abstract, Introduction, Conclusion
2. This ensures summary sections accurately reflect the full survey content

Summary Section Standards

Abstract (300-500 words):

• Continuous narrative, NO bullet points or bold labels
• Must cover: background → gap → scope → key findings → outlook

Introduction:

• Continuous narrative, NO subsections or bullet points
• Must cover: research background → why traditional methods fail → method summary → scope & organization

Conclusion:

• Summarize findings, state which paradigm is most promising
• Respond to user's original research goal
• Provide clear "next step" recommendation

Stage 6: Assemble Final Survey

Assemble sections in outline order, then append formatted references:

**1. Title** (Year). _Authors_. *Venue*. Citations: N. [[Link]](url)

Save to /artifacts/survey-{topic}-{date}.md.

Core Quality Principles

1. Build taxonomy, don't enumerate: Cluster papers by technical mechanism, not chronology. This is the defining characteristic of a survey vs. a summary.
2. Critical insight over description: For EVERY method, analyze WHY it works, WHAT trade-off it makes, WHERE it fails. This separates survey-grade writing from shallow summaries.
3. Goal-centric filtering: Every piece of information must answer "How does this help achieve the research goal?" Discard information that doesn't serve the goal, even if it's interesting.
4. Strict terminology fidelity: Use the user's exact technical terms. Do NOT drift to related but different concepts.
5. Dense citations: Ground ALL claims with numbered citations [X]. Nearly every sentence should reference at least one paper.
6. Zero vagueness: Replace generic statements with specific method names, dataset names, metric values, and problem descriptions.
7. Visual structure: Use Markdown tables extensively — paradigm comparison, intra-paradigm method comparison, benchmark tables, metric tables.

Reference Materials

Handoff