review-writing

name	review-writing
description	MANDATORY skill for ALL literature review writing. Use when user says '写综述', 'literature review', 'review writing', 'write review', or any request involving literature/paper review. This is the ONLY approved method for writing literature reviews in this project.
argument-hint	["topic or section reference from plan04.md"]
allowed-tools	Bash(*), Read, Write, Edit, Grep, Glob, WebSearch, WebFetch, Agent

Skill-5: Literature Review Writing

Write a comprehensive literature review for: $ARGUMENTS

Overview

This skill writes a structured, citation-rich literature review for a Graph RAG research paper. It follows a professional academic workflow: Zotero baseline → multi-database supplementation → thematic synthesis with figures and summary tables.

Key differentiator: Every review section includes a summary table (inspired by "Artificial intelligence and smart vision for building and construction 4.0") that synthesizes reviewed papers by method, application, innovation, KG content, etc.

Inputs

READ before starting:

1. output/ideas/idea01.md — research topic, innovation points, literature review framework (Section 8)
2. output/plans/plan04.md — Phase G tasks + review section structure (planned in idea/plan)
3. output/ontology/ontology03.md — domain concepts to guide search terms

The review section structure should already be defined in idea01.md Section 8 and plan04.md Phase G. If not, generate a structure and get user approval.

Outputs

output/papers/
├── literature_review.md             # Final review document
├── review_tables/                    # Summary tables per section (CSV)
│   ├── table_graphrag_methods.csv
│   ├── table_kg_construction.csv
│   ├── table_ontology_approaches.csv
│   └── ...
└── review_data/                      # Search & citation data
    ├── search_log.md                 # All search queries + results
    ├── paper_database.json           # All collected paper metadata
    ├── zotero_baseline.json          # Papers from Zotero
    └── citation_verification.json    # DOI verification results

output/figure_prompts/review/         # Figure prompts for skill-8
├── fig_prisma_flowchart.md          # PRISMA search→screen→include flow
├── fig_research_taxonomy.md         # Classification of reviewed papers
└── fig_publication_trend.md         # Publication year distribution (optional)

References & Scripts

This skill includes bundled reference materials:

Resource	Path	Purpose
Citation Styles	references/citation_styles.md	APA, Nature, Vancouver formatting
Database Strategies	references/database_strategies.md	Multi-database search guidance
Review Template	assets/review_template.md	Complete review document template
Search Script	scripts/search_databases.py	Aggregate and deduplicate results
Verify Script	scripts/verify_citations.py	DOI verification and metadata check

Workflow

Phase 1: Review Framework Confirmation

1. Read idea01.md Section 8 (Literature Review Framework) and plan04.md Phase G
2. Extract the planned review sections, e.g.:
   • 2.1 Knowledge Graphs in [Domain]
   • 2.2 Ontology Design for [Domain]
   • 2.3 Graph RAG and Retrieval-Augmented Generation
   • 2.4 Chatbot/QA Systems for [Domain]
   • 2.5 Research Gaps and Positioning
3. Present to user for confirmation. Wait for approval.

Phase 2: Zotero Baseline Collection

Start from the user's existing Zotero library to establish a baseline of known relevant papers.

1. Use Zotero MCP (preferred) to search the user's library:

   • Search by topic keywords related to each review section
   • Retrieve: title, authors, year, DOI, journal, abstract, tags
   • If Zotero MCP is not configured, fall back to manual export (see Introduction.md Section 8.5)

2. From Zotero results:

   • Build initial paper profile per paper
   • Classify each paper into review sections
   • Identify coverage gaps: which review sections have < 3 papers?
   • Save to output/papers/review_data/zotero_baseline.json

3. Report baseline to user:

   Zotero baseline: [N] papers loaded
   Coverage:
   - Section 2.1 (Knowledge Graphs): [N] papers
   - Section 2.2 (Ontology Design): [N] papers
   - Section 2.3 (Graph RAG): [N] papers ⚠️ needs more
   - Section 2.4 (Chatbot/QA): [N] papers ⚠️ needs more
   
   I'll supplement gaps via database search.
   

Phase 3: Multi-Database Supplementation

Fill gaps identified in Phase 2. Target: ~20 papers total across all sections.

3.1 Search Strategy

For each under-covered section, run targeted searches:

Semantic Scholar API (primary):

# WebFetch: https://api.semanticscholar.org/graph/v1/paper/search
# params: query, year (last 5 years), fields, limit
# Fields: title, abstract, authors, year, venue, externalIds, citationCount, isOpenAccess

arXiv API:

# WebFetch: http://export.arxiv.org/api/query
# search_query: "knowledge graph" AND "retrieval augmented generation"
# max_results: 20

Google Scholar (via WebSearch):

WebSearch: "graph RAG" "knowledge graph" site:scholar.google.com
WebSearch: "[domain] ontology knowledge graph" site:sciencedirect.com

3.2 Search Log

Document ALL searches in output/papers/review_data/search_log.md:

## Search Log

### Search 1: Semantic Scholar — Graph RAG
- **Date**: [date]
- **Query**: "graph RAG knowledge graph retrieval"
- **Filters**: year >= 2021, fields of study: Computer Science
- **Results**: [N] papers
- **Selected**: [N] papers (after title/abstract screening)

### Search 2: arXiv — Ontology Design
...

3.3 Paper Selection Criteria

• Recency: Last 5 years (2021-2026), prefer last 3 years
• Relevance: Directly related to review section topic
• Quality: Q1/Q2 journals preferred, top conferences accepted
• Citation count: Prioritize highly cited papers (see references/database_strategies.md)
• Diversity: Cover different approaches/methods, not just one group's work

3.4 Build Complete Paper Database

For each selected paper, collect:

{
  "id": "P01",
  "title": "...",
  "authors": "...",
  "year": 2024,
  "journal": "...",
  "doi": "10.xxxx/...",
  "abstract": "...",
  "citation_count": 45,
  "review_section": "2.3",
  "source": "semantic_scholar",
  "method": "Graph RAG with subgraph retrieval",
  "application": "Medical QA",
  "innovation": "Multi-hop reasoning over KG",
  "kg_content": "Disease-symptom-treatment graph",
  "limitations": "Small-scale evaluation"
}

Save to output/papers/review_data/paper_database.json.

Phase 4: Thematic Synthesis with Summary Tables

This is the core writing phase. For each review section:

4.1 Writing Pattern

For each section (e.g., "2.3 Graph RAG"):

1. Opening paragraph: Define the concept, why it matters, how it connects to the research
2. Thematic synthesis: Group papers by sub-theme, synthesize across papers (NOT study-by-study)
3. Summary table: A structured table summarizing all reviewed papers in this section
4. Gap identification: What's missing? What does this review section motivate?

4.2 Summary Table Design (Mandatory per section)

Inspired by "AI and smart vision for building and construction 4.0" paper.

Example for Graph RAG section:

**Table X.** Summary of Graph RAG approaches in recent literature.

| Ref | Authors (Year) | Method | Application Scenario | KG Content | Innovation | Evaluation | Limitations |
|-----|---------------|--------|---------------------|-----------|------------|------------|-------------|
| [1] | Zhang et al. (2024) | Subgraph retrieval + LLM | Medical QA | Disease-symptom-treatment | Multi-hop reasoning | ROUGE-L: 0.72 | Small dataset |
| [2] | Li et al. (2023) | Hybrid graph-vector RAG | Legal QA | Case-law graph | Citation chain traversal | Human eval: 4.2/5 | Domain-specific |
| ... | ... | ... | ... | ... | ... | ... | ... |

Table columns adapt per section:

Review Section	Recommended Table Columns
Knowledge Graphs in [Domain]	Authors, Year, KG Type, Domain, Nodes/Edges Scale, Construction Method, Application
Ontology Design	Authors, Year, Methodology, Domain, Concepts Count, Relations Count, Validation
Graph RAG	Authors, Year, Method, Application, KG Content, Innovation, Evaluation, Limitations
Chatbot/QA Systems	Authors, Year, Architecture, Domain, RAG Type, User Study, Performance

Also save each table as CSV to output/papers/review_tables/.

4.3 Figures (Mandatory)

At minimum, include:

1. PRISMA Flow Diagram (Mermaid): Show search → screening → inclusion pipeline
2. Research Taxonomy/Classification (Mermaid): Categorize reviewed papers by theme/method

Optional but encouraged: 3. Timeline/trend chart: Publication year distribution 4. Method comparison diagram: Visual comparison of approaches

graph TD
    A[Initial Search<br>N papers] --> B[Deduplication<br>N unique]
    B --> C[Title/Abstract Screening<br>N remaining]
    C --> D[Full Review<br>N included]
    D --> E1[Section 2.1: KG<br>N papers]
    D --> E2[Section 2.2: Ontology<br>N papers]
    D --> E3[Section 2.3: Graph RAG<br>N papers]
    D --> E4[Section 2.4: Chatbot/QA<br>N papers]

Phase 5: Citation Formatting & Verification

1. Choose citation style (default: numbered, e.g., [1], [2]):

   • Follow references/citation_styles.md
   • Typically Nature/Vancouver for engineering journals

2. In-text citations: Every claim must be cited. Example:

   Graph RAG has shown improved performance over traditional RAG in domain-specific QA tasks [3, 7, 12].
   

3. Verify all DOIs:

   python scripts/verify_citations.py output/papers/literature_review.md
   

4. Build reference list at the end of the review with complete metadata.

Phase 6: Assembly & Polish

Assemble the final output/papers/literature_review.md:

# Literature Review: [Title]

## 1. Introduction to the Review
[Why this review is needed, scope, methodology brief]

## 2. Review Methodology
### 2.1 Search Strategy
[Databases searched, keywords, date range]
### 2.2 Selection Criteria
[Inclusion/exclusion]
### 2.3 PRISMA Flow
[Mermaid diagram]

## 3. [Section Title from plan — e.g., Knowledge Graphs in Disaster Management]
[Thematic synthesis paragraphs with citations]

**Table 1.** Summary of knowledge graph approaches in [domain].
| Ref | Authors | Year | ... |

## 4. [Section Title — e.g., Ontology Design Methods]
[Thematic synthesis with citations]

**Table 2.** Comparison of ontology design methodologies.
| Ref | Authors | Year | Methodology | ... |

## 5. [Section Title — e.g., Graph RAG and Retrieval-Augmented Generation]
[Thematic synthesis with citations]

**Table 3.** Summary of Graph RAG approaches.
| Ref | Authors | Year | Method | Application | KG Content | Innovation | Evaluation | Limitations |

## 6. [Section Title — e.g., Chatbot and QA Systems]
[Thematic synthesis with citations]

**Table 4.** Comparison of chatbot/QA systems.
| Ref | Authors | Year | Architecture | ... |

## 7. Research Gaps and Positioning
[Gaps identified across all sections → how this paper addresses them]

**Figure X.** Research taxonomy and positioning of this work.
[Mermaid diagram showing where this paper fits]

## References
[1] Author A, Author B. Title. *Journal*. Year;Vol:Pages. doi:...
[2] ...

Phase 7: Update plan04.md

After completion, follow Auto-Update Protocol for Phase G tasks (G.1, G.2, G.3).

Key Rules

1. Zotero first, then supplement: Build on user's existing library, don't start from scratch.
2. ~20 papers total: Quality over quantity. Prefer highly cited Q1 papers.
3. Summary table per section: Mandatory. Columns adapted to section topic.
4. At least 2 figures: PRISMA flow + research taxonomy at minimum.
5. Every claim cited: No uncited statements in the review.
6. Thematic synthesis, not study-by-study: Group by theme, compare across papers.
7. Verify all DOIs: Run scripts/verify_citations.py before finalizing.
8. Review structure from plan: Follow the section structure defined in idea01.md / plan04.md.
9. Save all tables as CSV: For reuse in paper writing (skill-6).
10. Always update plan04.md after completing Phase G tasks.
11. Figure prompts: After writing the review, generate figure prompt files to output/figure_prompts/review/. Follow the format in Introduction.md Section 9.2. Minimum: PRISMA flowchart + research taxonomy.
12. Zotero MCP preferred: Use Zotero MCP tools to search the user's library. See Introduction.md Section 8 for configuration.

Composing with Other Skills

/idea-checking            → idea01.md (Section 8: review framework)
/research-planning        → plan04.md (Phase G: review tasks + section structure)
/review-writing           → literature_review.md + figures + tables  ← YOU ARE HERE
/paper-writing            → integrates literature_review.md as review section