Research Papers Skill

Open-source agent skill for researching, filtering, comparing, and synthesizing scientific papers across:

• arXiv
• OpenAlex
• ACM Digital Library
• DBLP
• Semantic Scholar
• PubMed
• SSRN

Works with Claude Code and Codex. This repository packages one skill: research-papers/.

What It Does

The skill helps an agent:

• turn vague requests such as "find the top papers on X" into a structured literature search
• choose the right source for the domain and task
• use official APIs where they exist
• keep preprints, working papers, reviews, and peer-reviewed work clearly separated
• rank papers by relevance, evidence quality, influence, and recency
• synthesize the results into shortlists, comparative reviews, evidence briefs, or literature maps

Install

Quick install (interactive)

git clone <this-repo>
cd skill-research-papers
bash scripts/install-local.sh

The script asks which platform you're using and copies the skill to the right location.

Quick install (non-interactive)

# Claude Code
bash scripts/install-local.sh --claude

# Codex
bash scripts/install-local.sh --codex

Manual install

Copy the research-papers/ folder into your platform's skills directory:

Platform	Destination
Claude Code	~/.claude/skills/research-papers
Codex	~/.codex/skills/research-papers

For Claude Code, you can rename the folder to research for a shorter /research command:

cp -R research-papers ~/.claude/skills/research

Use

Claude Code

The skill is invoked automatically when Claude detects a research need, or manually:

/research-papers

Example prompts:

• "Find the best survey papers on retrieval-augmented generation"
• "Compare seminal and recent graph neural network papers"
• "Build a reading list on causal representation learning for healthcare"

Codex

Use $research-papers to find the best survey papers on retrieval-augmented generation.
Use $research-papers to compare seminal and recent graph neural network papers.
Use $research-papers to build a reading list on causal representation learning for healthcare.

Environment Variables

Optional but recommended for better API reliability:

Variable	Used by	Purpose
OPENALEX_API_KEY	OpenAlex	Higher rate limits
SEMANTIC_SCHOLAR_API_KEY	Semantic Scholar	Reliable authenticated access
NCBI_API_KEY	PubMed	Higher rate limits
NCBI_EMAIL	PubMed	Good citizen identification
NCBI_TOOL	PubMed	Tool identification

Add them to your shell profile (e.g., ~/.zshrc or ~/.bashrc):

export OPENALEX_API_KEY="your-key-here"

Source Model

API-first sources:

• arXiv
• OpenAlex
• DBLP
• Semantic Scholar
• PubMed

Web-first sources:

• ACM Digital Library
• SSRN

Important nuance:

• Semantic Scholar is usable as an API source, but anonymous access can be unstable or rate-limited.
• DBLP author search tested cleanly, while a simple publication search returned a server-side 500 during live tests on March 8, 2026.
• ACM DL and SSRN should not be treated as stable public JSON API sources.

Verify

Run the bundled smoke tests from your own environment:

bash research-papers/scripts/smoke-test-sources.sh

Repository Layout

.
├── README.md
├── LICENSE
├── scripts/
│   └── install-local.sh
└── research-papers/
    ├── SKILL.md
    ├── agents/openai.yaml       # Codex only
    ├── references/
    │   ├── source-guides.md
    │   ├── api-playbook.md
    │   ├── screening-rubric.md
    │   ├── smoke-tests.md
    │   └── output-shapes.md
    └── scripts/
        └── smoke-test-sources.sh

The skill folder intentionally stays minimal. Public project documentation lives at the repository root, not inside the skill itself.

Platform Notes

Claude Code

• The SKILL.md frontmatter includes context: fork, agent: general-purpose, and model: opus — these configure Claude Code to run the skill as an isolated subagent using the Opus model.
• Claude Code ignores agents/openai.yaml.
• The install script removes the agents/ directory when installing for Claude Code.

Codex

• Codex ignores the Claude Code frontmatter fields (context, agent, model).
• Codex uses agents/openai.yaml for interface configuration.

License

MIT. See LICENSE.