Survey Generation

Survey Generation

Generate complete academic survey papers with structured outline, RAG-based writing, and citation validation.

Input

• $0 — Survey topic or research area

Scripts

Literature search

python ~/.claude/skills/deep-research/scripts/search_semantic_scholar.py \
  --query "relevant search query" --max-results 50

References

• Survey prompts (outline, writing, citation, coherence): ~/.claude/skills/survey-generation/references/survey-prompts.md

Workflow (from AutoSurvey)

Step 1: Collect Papers

1. Search Semantic Scholar / arXiv for papers on the topic
2. Collect 50-200 relevant papers with titles and abstracts
3. Filter by relevance and citation count

Step 2: Generate Outline (Multi-LLM Parallel)

1. Generate N rough outlines independently (parallel)
2. Merge outlines into a single comprehensive outline
3. Expand each section into subsections
4. Edit final outline to remove redundancies

Step 3: Write Subsections (RAG-Based)

For each subsection:

1. Retrieve relevant papers for the subsection topic
2. Generate content with inline citations [paper_title]
3. Enforce minimum word count per subsection
4. Only cite papers from the provided list

Step 4: Validate Citations

For each subsection:

1. Check that cited paper titles are correct
2. Verify citations support the claims made
3. Remove or correct unsupported citations
4. Use NLI (Natural Language Inference) for claim-source faithfulness

Step 5: Enhance Local Coherence

For each subsection:

1. Read previous and following subsections
2. Refine transitions and flow
3. Preserve core content and citations
4. Ensure smooth reading experience

Step 6: Convert Citations to BibTeX

1. Replace [paper_title] with \cite{key}
2. Generate BibTeX entries for all cited papers
3. Validate all citation keys exist in .bib file

Output Structure

survey/
├── main.tex          # Complete survey paper
├── references.bib    # All citations
├── outline.json      # Survey outline
└── sections/         # Individual section files

Rules

• Only cite papers from the collected paper list — never hallucinate citations
• Each subsection must meet minimum word count
• No duplicate subsections across sections
• Citation validation is mandatory before final output
• Local coherence enhancement must preserve all citations
• The survey should be comprehensive and logically organized

Related Skills

• Upstream: [deep-research](../deep-research/), [literature-search](../literature-search/), [literature-review](../literature-review/)
• See also: [related-work-writing](../related-work-writing/)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ARAVINDAN20
• Source: ARAVINDAN20/Claude-Research-Paper-OS
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.