convez.ai
Skill
chenlu-hung

literature-review

chenlu-hung/literature-review

Ver no GitHub
InicialPluginScore · de 0 a 100

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-review

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

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:

  1. Scoping — Frame the question with PMDB (Problem / Method / Data / Baseline) instead of clinical PICO
  2. Multi-source search — arXiv + Semantic Scholar + OpenAlex + CrossRef (all free, no API key)
  3. Screening + deduplication — Title → abstract → full-text with PRISMA flow tracking
  4. Data extraction — ML-specific schema capturing method, dataset, code link, compute, seeds
  5. Quality assessment — 6-dimensional reproducibility score (0–18) per paper
  6. 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
  7. Citation verification — Every arXiv ID and DOI is resolved against arXiv/CrossRef; mismatched metadata flagged
  8. PDF generation — pandoc + xelatex with APA / Nature / IEEE / ACM / Vancouver styles

Differences from generic literature-review skills

AspectGeneric biomed skillliterature-review-ml
Scoping framePICO (clinical)PMDB (Problem/Method/Data/Baseline)
Primary search sourcePubMedarXiv + Semantic Scholar
Quality assessmentCochrane RoB / AMSTAR 2NeurIPS/JMLR reproducibility checklist
Citation impact metricRaw citation countinfluentialCitationCount (Semantic Scholar)
Synthesis outputThematic prose onlyThematic prose + benchmark comparison table
Citation verificationDOI resolvesDOI + arXiv ID + metadata consistency check
DependenciesPaid APIs sometimes requiredAll 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 source
  • screened/ — deduplicated candidates with screening decisions
  • extracted/ — per-paper YAML extracted using the ML-specific schema
  • review.md + review.pdf — final output
  • verification_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.

Ver o original no GitHub

Bom para

Relacionados