literature-review
chenlu-hung/literature-review
Claude Code skill for systematic literature reviews in ML and statistics — arXiv/Semantic Scholar/OpenAlex search, PMDB scoping, reproducibility scoring, citation verification, PDF output.
Instalação
npx convez-ai chenlu-hung/literature-review1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
literature-review-ml
A Claude Code skill for conducting systematic literature reviews tailored to machine learning and statistics research.
Designed for survey-paper rigor — every paper passes through a documented PRISMA-adjacent flow, all citations are arXiv-ID / DOI verified, and reproducibility is explicitly assessed using a NeurIPS/JMLR-style checklist.
What it does
Given a research topic, this skill guides Claude through 8 phases:
- Scoping — Frame the question with PMDB (Problem / Method / Data / Baseline) instead of clinical PICO
- Multi-source search — arXiv + Semantic Scholar + OpenAlex + CrossRef (all free, no API key)
- Screening + deduplication — Title → abstract → full-text with PRISMA flow tracking
- Data extraction — ML-specific schema capturing method, dataset, code link, compute, seeds
- Quality assessment — 6-dimensional reproducibility score (0–18) per paper
- Thematic synthesis — Organized by method family, not paper-by-paper; benchmark comparison table; AI-tell prose pass so the writing doesn't read as machine-generated
- Citation verification — Every arXiv ID and DOI is resolved against arXiv/CrossRef; mismatched metadata flagged
- PDF generation — pandoc + xelatex with APA / Nature / IEEE / ACM / Vancouver styles
Differences from generic literature-review skills
| Aspect | Generic biomed skill | literature-review-ml |
|---|---|---|
| Scoping frame | PICO (clinical) | PMDB (Problem/Method/Data/Baseline) |
| Primary search source | PubMed | arXiv + Semantic Scholar |
| Quality assessment | Cochrane RoB / AMSTAR 2 | NeurIPS/JMLR reproducibility checklist |
| Citation impact metric | Raw citation count | influentialCitationCount (Semantic Scholar) |
| Synthesis output | Thematic prose only | Thematic prose + benchmark comparison table |
| Citation verification | DOI resolves | DOI + arXiv ID + metadata consistency check |
| Dependencies | Paid APIs sometimes required | All free, no API key required |
Installation
git clone https://github.com/chenlu-hung/literature-review ~/.claude/skills/literature-review-ml
Then verify:
ls ~/.claude/skills/literature-review-ml/SKILL.md
In Claude Code, the skill becomes available as /literature-review-ml. The description field in SKILL.md also lets Claude auto-invoke it when you ask for an ML/stats literature review.
Dependencies
Python (standard library only, except pyyaml for the extraction template):
pip install pyyaml
For PDF output:
# macOS
brew install pandoc
brew install --cask mactex
# Linux
apt-get install pandoc texlive-xetex
Check with:
python ~/.claude/skills/literature-review-ml/scripts/generate_pdf.py --check-deps
Quick usage
In Claude Code:
/literature-review-ml neural operators for parametric PDEs
Or invoke directly via natural language:
Conduct a systematic literature review on operator learning methods for parametric PDEs, focusing on FNO and DeepONet variants since 2020.
Claude will then walk through all 8 phases, producing:
sources/— JSON from each search sourcescreened/— deduplicated candidates with screening decisionsextracted/— per-paper YAML extracted using the ML-specific schemareview.md+review.pdf— final outputverification_report.json— citation verification results
A complete example workflow is included in SKILL.md under "Example: Operator Learning for PDEs Review".
Standalone script usage
The Python scripts also work as standalone CLIs:
# Search arXiv
python scripts/search_arxiv.py "neural operator" \
--categories cs.LG,stat.ML,math.NA \
--year-start 2020 --max-results 100 \
--output sources/arxiv.json
# Search Semantic Scholar (for citation counts)
python scripts/search_semantic_scholar.py "neural operator parametric PDE" \
--min-citations 10 --max-results 50 \
--output sources/s2.json
# Aggregate + deduplicate + rank
python scripts/aggregate.py sources/*.json \
--deduplicate --rank influential_citations \
--format markdown --output screened/candidates.md
# Verify citations in your draft
python scripts/verify_citations.py review.md \
--output verification_report.json
# Generate PDF
python scripts/generate_pdf.py review.md \
--citation-style nature \
--output review.pdf
Repository layout
literature-review-ml/
├── SKILL.md # main skill instructions (loaded by Claude)
├── assets/
│ ├── review_template.md # full survey-paper skeleton
│ └── extraction_template.yaml # per-paper extraction schema
├── references/
│ ├── ml_venues.md # ML / stats / SciML venue tiers
│ ├── search_strategies.md # arXiv categories, S2 / OpenAlex tips
│ ├── citation_styles.md # APA / Nature / IEEE / ACM / Vancouver
│ └── prose_hygiene.md # AI-tell checklist for survey prose (stop-slop, adapted)
└── scripts/
├── search_arxiv.py
├── search_semantic_scholar.py
├── search_openalex.py
├── search_crossref.py
├── aggregate.py
├── verify_citations.py
└── generate_pdf.py
Acknowledgments
This skill draws workflow inspiration from K-Dense-AI/scientific-agent-skills (MIT-licensed). The PMDB frame, ML/stats venue tiers, reproducibility scoring rubric, and metadata-consistency citation verification are original to this skill.
License
MIT — see LICENSE.