Paper Finder

Systematic paper discovery and prioritization for research projects.

When to Use

• Starting a literature review
• Looking for related work on a topic
• Finding baseline papers for experiments
• Searching for methodology references
• Building a citation list for a research paper

How to Use

Run the helper script from your workspace:

{{ research_python_path }} .claude/skills/paper-finder/scripts/find_papers.py "your research topic"

Options:

• --mode fast (default): Quick search
• --mode diligent: Thorough search (recommended for comprehensive review)
• --format json: Output as JSON instead of text

Example:

{{ research_python_path }} .claude/skills/paper-finder/scripts/find_papers.py "hypothesis generation with large language models" --mode fast

Search Strategy

Query Formulation

Use structured queries for better results:

1. Core concept + Method: "transformer attention mechanism"
2. Problem + Domain: "few-shot learning natural language processing"
3. Technique + Application: "graph neural networks drug discovery"

Multi-Stage Search

1. Broad scan: Start with general topic terms
2. Focused dive: Use specific method/technique names from initial results
3. Citation chase: Search for highly-cited papers referenced in relevant work

Citation Prioritization

Relevance Tiers

Relevance Score	Priority	Action
3 (High)	Must read	Download PDF, read fully, cite
2 (Medium)	Should read	Read abstract + intro, cite if relevant
1 (Low)	Optional	Skim abstract, cite only if needed
0 (Not relevant)	Skip	Do not include

Citation Count Thresholds

Category	Citation Count	Interpretation
Seminal	1000+	Foundational work, must cite
Well-established	100-999	Widely accepted, cite if relevant
Recent/Emerging	10-99	Current research, cite for novelty
New	<10	Very recent, check publication venue

Venue Tiers (ML/AI Focus)

Tier 1 (Top venues, high credibility):

• NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR, ICCV

Tier 2 (Strong venues):

• AAAI, IJCAI, NAACL, COLING, ECCV, WACV

Tier 3 (Good venues):

• *ACL workshops, COLM, EACL, CoNLL

Preprints (arXiv):

• Check for peer-reviewed version first
• Cite arXiv only if no published version exists

Screening Workflow

Phase 1: Title Screening

• Review titles from search results
• Mark papers as "include", "exclude", or "maybe"
• Goal: ~50% reduction

Phase 2: Abstract Screening

• Read abstracts for included/maybe papers
• Evaluate: relevance, methodology, findings
• Goal: Identify key papers for deeper reading

Phase 3: Full-Text Review

• Download and read full PDFs for key papers
• Extract: methods, results, limitations, citations
• Use the PDF chunker for detailed reading (see below)

Output Structure

Returns relevance-ranked papers with:

• Title, authors, year
• Abstract (already extracted)
• URL for download
• Relevance score (0-3, focus on papers with score >= 2)
• Citation count

After Finding Papers

1. Download PDFs for papers with relevance >= 2
2. Read abstracts first (already provided in output)
3. Only read full PDFs for most relevant papers
4. Write notes to literature_review.md immediately
5. Track citations for references.bib

Reading Large PDFs

Use the PDF chunker to split papers into smaller PDF files that can be read directly. This preserves all formatting perfectly (unlike text extraction which loses formatting).

Dependencies:

# Using uv (recommended):
uv add --project {{ research_env_dir }} pypdf

How to run:

{{ research_python_path }} .claude/skills/paper-finder/scripts/pdf_chunker.py <pdf_path>

Options:

• --pages-per-chunk N: Number of pages per chunk (default: 1)
• --output-dir DIR: Output directory (default: <pdf_dir>/pages)

Output:

• Creates PDF chunk files: <pdf_name>_chunk_001.pdf, <pdf_name>_chunk_002.pdf, etc.
• Creates a manifest: <pdf_name>_manifest.txt listing all chunks with page ranges

Integration with screening workflow:

1. Run the chunker on papers before detailed reading
2. For abstract skimming: read only chunk 1 (page 1 or pages 1-3)
3. For deep reading: read ALL chunk PDFs sequentially, writing notes after each
4. Check the manifest to see how many chunks exist
5. IMPORTANT: Do not skip chunks - methodology and results are in later chunks

If Paper-Finder Service Not Running

The script will show a fallback message. Use manual search instead:

• arXiv: https://arxiv.org
• Semantic Scholar: https://www.semanticscholar.org
• Papers with Code: https://paperswithcode.com
• Google Scholar: https://scholar.google.com

Manual search works well - paper-finder is just a convenience for faster, more targeted results.

References

See references/ folder for:

• search_strategies.md: Detailed search query formulation
• prioritization_guide.md: Extended prioritization criteria