Survey Writer Skill

Orchestrate end-to-end academic survey writing with MCP-based paper retrieval, parallel subagent analysis, and iterative synthesis.

Workflow

[Input: Topic / Research Question]
      ↓
[Phase 1] Topic Scoping
      ↓
[Phase 2] Paper Discovery (ArXiv MCP)
      ↓
[Phase 3] Parallel Analysis (Task Subagents)
      ↓
[Phase 4] Survey Writing (Iterative)
      ↓
[Phase 5] Verification & Cross-check
      ↓
[Output: survey.md + references.bib]

Phase 1: Topic Scoping

Accept the survey topic from the user and define the scope:

1. •Define the topic clearly (e.g., "AI Methods for Playing Othello")
2. •Identify subtopics to cover (e.g., classical methods, MCTS, deep learning, RL)
3. •Generate search queries — create 6-10 diverse queries covering:
   • •Core topic terms
   • •Alternate names or synonyms
   • •Specific method families (RL, MCTS, neural networks, evolutionary)
   • •Broader related domains
4. •Set inclusion criteria — arXiv papers, conference papers, relevance threshold

Phase 2: Paper Discovery

Use mcp__arxiv-mcp-server__search_papers to find candidate papers:

For each search query:
  → mcp__arxiv-mcp-server__search_papers(query, max_results=10)
  → Collect paper IDs, titles, abstracts
  → Deduplicate across queries
  → Filter to 10-15 most relevant papers

Search strategy:

• •Run multiple queries to maximize coverage across subtopics
• •Prioritize papers that directly address the survey topic
• •Include foundational papers and recent advances
• •Ensure diversity of methods (classical, learning-based, hybrid)

Phase 3: Parallel Analysis with Task Subagents

Spawn Task tool subagents in parallel batches of 3-4 papers each.

Each subagent performs:

1. •Download paper via mcp__arxiv-mcp-server__download_paper(paper_id)
2. •Read content via mcp__arxiv-mcp-server__read_paper(paper_id)
3. •Extract structured information:
   • •Title, authors, year, arXiv ID
   • •Research question / motivation
   • •Method summary
   • •Key results and metrics
   • •Strengths and limitations
   • •Relevance to survey topic
4. •Generate BibTeX entry with real metadata from the paper

Subagent prompt template:

Download and analyze arXiv paper {paper_id}.
Use mcp__arxiv-mcp-server__download_paper to download, then
mcp__arxiv-mcp-server__read_paper to read the full text.
Extract: title, authors, year, research question, method,
results, strengths, limitations. Generate a BibTeX entry.
Return all as structured text.

Batching strategy:

• •Batch 1: Papers 1-4 (launch in parallel)
• •Batch 2: Papers 5-8 (launch in parallel)
• •Batch 3: Papers 9-12 (launch in parallel)
• •Collect all results before proceeding to Phase 4

Phase 4: Survey Writing

Synthesize all paper analyses into a structured survey.md:

Document structure:

1. •Introduction — domain motivation, scope, contributions
2. •Background — foundational concepts, problem formulation
3. •Thematic sections (3-5) — grouped by methodology family
4. •Discussion — cross-cutting themes, comparative analysis, open problems
5. •Conclusion — summary of findings, future directions
6. •References — cite all analyzed papers

Writing principles:

• •Topic-first paragraphs: Lead with the main point, then support
• •Prose format: No bullet points in the body; write flowing paragraphs
• •Inline citations: Use [@citekey] format throughout
• •Critical analysis: Don't just describe — compare, contrast, evaluate
• •Comparative tables: Include method comparison tables where appropriate
• •Concrete numbers: Quote specific results with sources
• •Logical flow: Each section builds on the previous one

Iterative improvement:

• •Write first draft focusing on completeness
• •Review for logical flow and coherence
• •Ensure every cited paper has a matching BibTeX entry
• •Verify no hallucinated claims or papers

Phase 5: Verification

Run final quality checks before delivering:

• • references.bib has 10+ entries with real arXiv IDs
• • Every [@citekey] in survey.md has a matching references.bib entry
• • survey.md has introduction, organized body sections, and conclusion
• • Each paper receives critical analysis (not just description)
• • No hallucinated papers — all fetched via MCP tools
• • Comparative table or summary comparing methods
• • Research gaps and future directions discussed

If any check fails: Fix the issue and re-verify before delivering.

Output Files

MCP Tools Used

Quality Standards

• •All papers must be real, verifiable arXiv publications
• •Survey must demonstrate synthesis, not just paper-by-paper summaries
• •Critical analysis should identify strengths, limitations, and research gaps
• •Writing should be accessible to researchers familiar with AI but not necessarily the specific subfield