agent-research-skills

31 skills for Claude Code covering the full academic research paper lifecycle — from literature search to slide generation — plus GitHub repository analysis for research topics.

Extracted from 17 GitHub repos studying LLM-agent-driven research automation. See SKILLS_DESIGN.md for the original design specifications.

Installation

One-line install

npx skills add lingzhi227/agent-research-skills -g -a claude-code

Use the -g (global) flag. Scripts use ~/.claude/skills/ paths that require global installation.

Post-install setup

git clone https://github.com/lingzhi227/agent-research-skills.git /tmp/agent-research-skills
/tmp/agent-research-skills/install.sh
rm -rf /tmp/agent-research-skills

This installs slash commands, checks Python dependencies, and verifies script syntax.

Optional dependencies

Package	Required by	Install
Python 3	All scripts	brew install python3
PyMuPDF	self-review, deep-research (PDF parsing)	pip install PyMuPDF
numpy + scipy	data-analysis (statistical tests)	pip install numpy scipy

Optional configuration

1. Semantic Scholar API key (higher rate limits for literature search):

   • Get one at https://www.semanticscholar.org/product/api#api-key
   • Save in ~/keys.md: S2_API_Key: your-key-here

2. Output directory: Deep research outputs go to ~/deep-research-output/ by default.

Available Skills (31)

Phase 0: Research Discovery & Planning

Skill	Description	Scripts
github-research	6-phase GitHub repo discovery, analysis, and integration planning for research topics	13 scripts: search, clone, analyze structure/deps/impls, compare, compile report
deep-research	6-phase systematic literature survey (frontier → survey → deep dive → code → synthesis → report)	7 scripts: search APIs, PDF extraction, paper DB, BibTeX, report compilation
literature-search	Multi-source academic search (Semantic Scholar, arXiv, OpenAlex, CrossRef) with ranking	4 scripts: search_crossref.py, download_arxiv_source.py, search_openalex.py + shared
literature-review	Multi-perspective dialogue simulation with expert personas for grounded literature review	Shares search scripts
idea-generation	Generate and score research ideas (Interestingness/Feasibility/Novelty) with iterative refinement	1 script: novelty_check.py
novelty-assessment	Harsh-critic novelty evaluation with up to 10 rounds of literature search	Shares search scripts
research-planning	4-stage research plan design with task dependency graphs	Prompt-only

Phase 1: Method Design

Skill	Description	Scripts
atomic-decomposition	Decompose ideas into atomic concepts with bidirectional math ↔ code mapping	Prompt-only
algorithm-design	Algorithm pseudocode (LaTeX) + UML diagrams (Mermaid) with consistency verification	Prompt-only
math-reasoning	Derivations, proofs, formalization, statistical test selection, notation tables	Prompt-only
symbolic-equation	LLM-guided evolutionary search for scientific equation discovery	Prompt-only

Phase 2: Experiment Pipeline

Skill	Description	Scripts
experiment-design	4-stage progressive experiment planning (implement → tune → research → ablate)	1 script: design_experiments.py
experiment-code	ML training/evaluation pipeline generation with iterative improvement	Prompt-only
code-debugging	Structured error analysis with categorization and 4-retry fix loop	Prompt-only
data-analysis	Statistical analysis with 4-round code review and appropriate test selection	2 scripts: stat_summary.py, format_pvalue.py

Phase 3: Paper Writing

Skill	Description	Scripts
paper-writing-section	Section-by-section writing with section-specific guidance and two-pass refinement	Prompt-only
related-work-writing	Related Work section with thematic organization and compare-and-contrast style	Prompt-only
survey-generation	Complete survey paper via RAG-based subsection writing with citation validation	Shares search scripts
paper-to-code	Convert paper PDF to runnable code repo (Planning → Analysis → Coding pipeline)	Prompt-only

Phase 4: Figures, Tables & Citations

Skill	Description	Scripts
figure-generation	Publication-quality matplotlib figures with VLM feedback loop (10 figure types)	1 script: figure_template.py
table-generation	JSON/CSV → LaTeX booktabs tables with bold-best, significance stars, multi-dataset	1 script: results_to_table.py
citation-management	BibTeX harvesting, validation, deduplication, and auto-fix	2 scripts: harvest_citations.py, validate_citations.py
backward-traceability	Every PDF number hyperlinks to the code line that produced it	1 script: ref_numeric_values.py

Phase 5: LaTeX & Compilation

Skill	Description	Scripts
latex-formatting	Conference templates (ICML/ICLR/NeurIPS/AAAI/ACL), formatting fixes, pre-submission checks	2 scripts: latex_checker.py, clean_latex.py
paper-compilation	Full pdflatex+bibtex pipeline with auto-fix error correction loop	2 scripts: compile_paper.py, fix_latex_errors.py
excalidraw-skill	Programmatic Excalidraw diagramming via MCP tools with quality verification	MCP server (7 CJS files)

Phase 6: Review & Polish

Skill	Description	Scripts
self-review	3-persona automated review (NeurIPS form) with reflection and meta-review	2 scripts: extract_pdf_text.py, parse_pdf_sections.py
paper-revision	Map reviewer concerns to sections, apply targeted edits, verify improvements	Prompt-only
rebuttal-writing	Point-by-point rebuttal with evidence-based responses	Prompt-only
slide-generation	Paper → Beamer slides (extract elements, generate skeleton, simplify)	1 script: extract_paper_elements.py
paper-assembly	End-to-end pipeline orchestrator with 9-phase checkpointing	1 script: assembly_checker.py

Usage

Slash commands

/research transformer architectures for long-context reasoning

Natural language (skills activate automatically)

"Analyze GitHub repos for multi-agent coordination research"
"Do a literature review on protein folding with LLMs"
"Write the Methods section of my paper"
"Generate a comparison table from results.json"
"Review my paper draft before submission"
"Make slides from my paper"
"Check if my idea is novel"
"Design experiments for my contrastive learning method"

Direct script usage

# Search GitHub repos for a research topic
python ~/.claude/skills/github-research/scripts/search_github.py --query "multi-agent LLM coordination" --max-results 50 --output repos.jsonl

# Analyze a cloned repo's structure
python ~/.claude/skills/github-research/scripts/analyze_repo_structure.py --repo-dir ./my-repo --output analysis.json

# Search literature
python ~/.claude/skills/literature-search/scripts/search_crossref.py --query "attention mechanism" --rows 10

# Validate citations
python ~/.claude/skills/citation-management/scripts/validate_citations.py --tex paper/main.tex --bib paper/references.bib

# Generate experiment design
python ~/.claude/skills/experiment-design/scripts/design_experiments.py --method "contrastive learning" --task classification --format markdown

# Format p-values
python ~/.claude/skills/data-analysis/scripts/format_pvalue.py --values "0.001 0.05 0.23" --format latex

# Extract paper elements for slides
python ~/.claude/skills/slide-generation/scripts/extract_paper_elements.py --tex main.tex --output slides.tex

# Check paper pipeline completeness
python ~/.claude/skills/paper-assembly/scripts/assembly_checker.py --dir paper/ --verbose

Architecture

All skills follow the same structure:

skills/<skill-name>/
├── SKILL.md           # Skill definition (prompt, workflow, rules)
├── scripts/           # Executable tools (optional)
│   └── *.py           # CLI scripts with argparse, docstring headers
└── references/        # Reference docs (optional)
    └── *.md           # Templates, API docs, patterns

Design principles:

• Scripts are stdlib-only where possible (no heavy dependencies)
• Every script has --help, docstring header with usage examples, and argparse CLI
• Skills link to each other via ## Related Skills sections (upstream/downstream/see-also)
• Prompt-only skills are fully self-contained in SKILL.md

Requirements

• Claude Code
• Python 3
• Optional: PyMuPDF, numpy, scipy (see installation)