Research

Academic Paper Research & Dataset Discovery

Instructions and tools for searching, fetching, and extracting ML/AI research papers, plus finding and downloading datasets.

Prerequisites

which pdftotext || echo "MISSING: brew install poppler (macOS) or apt install poppler-utils (Linux)"

Available sources (7 total)

Source	Search	Fetch	Best for
arXiv	yes	yes	ML/AI preprints
Semantic Scholar	yes	yes	Citations, open-access PDFs
Papers with Code	yes	yes	Papers linked to GitHub repos
Hugging Face	yes	via arXiv	Trending daily papers
JMLR	yes	yes	Peer-reviewed ML journal
ACL Anthology	no	by ID	NLP conference papers
OpenScholar	no	no	Q&A synthesis over 45M papers (URL only)

Commands

Search

uv run ${CLAUDE_SKILL_DIR}/scripts/search.py "$ARGUMENTS" --source arxiv --max 5

Flags: --source (arxiv, semantic_scholar, papers_with_code, huggingface, jmlr, openscholar), --max N, --cat cs.AI,cs.LG, --sort relevance|date

Fetch metadata

uv run ${CLAUDE_SKILL_DIR}/scripts/fetch.py <paper_id>

Auto-detects source: 2401.12345 (arXiv), 2022.acl-long.220 (ACL), v22/19-920 (JMLR), 40-char hex (Semantic Scholar)

Download PDF

uv run ${CLAUDE_SKILL_DIR}/scripts/download.py <paper_id> -o ./papers/

Extract text

uv run ${CLAUDE_SKILL_DIR}/scripts/extract.py ./papers/<file>.pdf --max-pages 20

Guidelines

• Search 3-5 papers, not 20. Read abstracts first.
• Only download PDFs for papers the user cares about.
• Use --max-pages 10 for summaries; go deeper only when asked.
• Summarize key findings, then offer to dive deeper.
• Always cite: title, authors, year, source URL.

Additional tools

Multi-source concurrent search (alternative)

uv run ${CLAUDE_SKILL_DIR}/scripts/scientific_search.py "$ARGUMENTS" --max 10

Searches arXiv + Semantic Scholar concurrently with deduplication. Add --datasets for Kaggle/HuggingFace datasets.

Document analysis

uv run ${CLAUDE_SKILL_DIR}/scripts/analyze_document.py <url_or_path> [--max-pages 10] [--json]

Extracts text from PDFs, Word docs, text files. Supports URLs and local paths.

Dataset Discovery & Download

Available dataset sources (5 total, all free, no API keys)

Source	Search	Info	Download	Best for
HuggingFace	yes	yes	yes (parquet)	NLP, vision, audio datasets
OpenML	yes	yes	yes (ARFF/CSV)	Tabular/benchmark datasets
UCI	yes	yes	yes (CSV/ZIP)	Classic ML datasets
Papers with Code	yes	yes	no (links only)	Datasets linked to papers
Kaggle	yes	no	no (use kaggle CLI)	Competition & community datasets

Search datasets

uv run ${CLAUDE_SKILL_DIR}/scripts/datasets.py search "$ARGUMENTS" --source huggingface --limit 10

Flags: --source (huggingface, openml, uci, paperswithcode, kaggle), --limit N

Get dataset info

uv run ${CLAUDE_SKILL_DIR}/scripts/datasets.py info <dataset_id> --source huggingface

Shows description, columns, splits, download URLs. Works with: huggingface, openml, uci, paperswithcode.

Download dataset

uv run ${CLAUDE_SKILL_DIR}/scripts/datasets.py download <dataset_id> --source huggingface --output ./datasets --split train

Downloads to ./datasets/ directory. Flags: --output DIR, --split train|test|validation (HuggingFace only).

Dataset guidelines

• Search 3-5 datasets, compare sizes and licenses before downloading.
• Use info to check columns/splits before downloading.
• HuggingFace downloads as Parquet (read with pd.read_parquet()).
• OpenML downloads as ARFF, auto-converts to CSV when possible.
• For Kaggle downloads, use: kaggle datasets download -d <owner/dataset>

Rate limits (enforced automatically)

arXiv: 3s, Semantic Scholar: 4s, Papers with Code: 3s, JMLR: 3s per volume, OpenML: 2s, UCI: 2s, Kaggle: 2s.

Reference

• references/sources.md — API details, endpoints, rate limits for all sources

Paper Review

Structured framework for reviewing ML/AI research papers. Produces fair, constructive, conference-quality reviews.

Obtain the paper

Use the research scripts to get the paper content:

# Download by arXiv ID
uv run ${CLAUDE_SKILL_DIR}/scripts/download.py 2401.12345 --output ./papers

# Extract text from PDF
uv run ${CLAUDE_SKILL_DIR}/scripts/extract.py ./papers/2401.12345.pdf --max-pages 30

# Fetch metadata
uv run ${CLAUDE_SKILL_DIR}/scripts/fetch.py 2401.12345

If the user provides only a topic, search first, then review the selected paper.

Review template

Summary

2-3 sentences: What is the paper about? What is the key contribution?

Strengths

Evaluate each dimension:

• Novelty: Is the approach new? Does it advance the field?
• Experiments: Well-designed? Sufficient baselines?
• Clarity: Well-written? Easy to follow?
• Significance: Would this matter if results hold?

Weaknesses

Identify specific issues:

• Missing baselines or comparisons
• Claims not supported by evidence
• Methodology gaps or questionable choices
• Limited evaluation scope
• Scalability concerns

Methodology assessment

Dimension	Assessment	Notes
Splits	proper / questionable / missing	Train/val/test separation
Baselines	fair / unfair / missing	SOTA included?
Metrics	appropriate / limited / wrong	Multiple metrics?
Significance	reported / missing	Error bars, CIs, p-values
Ablations	thorough / partial / none	Component contributions

Reproducibility checklist

• Code available (or promised)?
• Dataset available or described sufficiently?
• Hyperparameters fully specified?
• Compute requirements stated (GPU type, hours)?
• Random seeds reported?
• Training details sufficient to reproduce?
• Preprocessing steps documented?

Questions for authors

3-5 specific questions that would strengthen the paper or clarify ambiguities.

Overall assessment

	Rating
Recommendation	Accept / Weak Accept / Borderline / Weak Reject / Reject
Confidence	High / Medium / Low
Impact	What would this enable if results hold?

Review ethics

• Be constructive — suggest improvements, don't just criticize
• Separate factual issues from opinions
• Acknowledge uncertainty in your assessment
• Evaluate what was claimed, not what wasn't attempted
• Consider the target venue and its standards
• Give credit where due — acknowledge genuine contributions

Comparative review (multiple papers)

When reviewing multiple papers on the same topic:

Dimension	Paper A	Paper B	Paper C
Method
Dataset
Best metric
Reproducibility
Novelty

Rank by overall contribution, noting complementary strengths.

---

Prototyping: Paper → Code

Convert a research paper, article, or technical document into a complete working code project.

uv run ${CLAUDE_SKILL_DIR}/scripts/prototype/main.py <source> -o ./prototype [-l python] [-v]

Argument	Description
<source>	PDF file, URL, .ipynb, .md, or .txt
-o	Output directory (default: ./prototype)
-l	Language: python, javascript, typescript, rust, go
-v	Verbose output

Pipeline

1. Extract — parse PDF, web page, notebook, or markdown
2. Analyze — detect algorithms, architectures, domain, dependencies
3. Select language — explicit flag → code in source → domain default → Python
4. Generate — complete project scaffold with algorithm implementations

Every generated project includes: main implementation, dependency manifest, test file, README, .gitignore.

Language defaults by domain

Domain	Default
ML / data science	Python
Web / frontend	TypeScript
Systems / performance	Rust
Explicit numpy/torch mention	Python
Explicit react/vue mention	JavaScript

Code quality requirements

• No TODOs or placeholders — all code must be complete and runnable
• Type hints on all functions (Python, TypeScript)
• Docstrings on all public functions
• Specific error handling (not bare except)
• At least one test file
• README includes: title, install, usage, source attribution

See references/prototype/generation-rules.md for full generation rules.