A 50+ agent skill suite for Claude Code that automates the full academic research pipeline — from paper discovery with citation verification to publication-ready manuscripts with simulated peer review.
Core guarantee: every citation is verified against its actual arXiv page. Unconfirmed papers are discarded. Zero hallucinated references.
LLMs hallucinate citations. This project solves that by building a multi-phase verification pipeline on top of Claude Code's skill system.
The suite covers five stages of academic research:
| Skill | Version | Agents | Scope |
|---|---|---|---|
| Discovery | v2.1 | 4-phase | Paper discovery via Semantic Scholar + arXiv APIs, community signal aggregation, citation graph expansion, importance scoring |
| Deep Research | v2.4 | 13 | Research question formulation (FINER), systematic literature search, cross-source synthesis, bias assessment, meta-analysis |
| Academic Paper | v2.4 | 12 | Structure architecture, argument building, draft writing, citation compliance (APA/IEEE/Chicago/MLA/Vancouver), LaTeX/PDF output |
| Paper Reviewer | v1.4 | 5 | Simulated peer review panel — Editor-in-Chief + 3 domain reviewers + Devil's Advocate, 0-100 rubric scoring |
| Pipeline | v2.8 | orchestrator | 11-stage workflow with mandatory integrity verification, two-round peer review, Socratic revision coaching |
The full pipeline produces 50-100 verified papers per topic, catches fabricated references at two verification checkpoints (Stage 2.5 and 4.5), and outputs publication-ready manuscripts.
git clone https://github.com/Leooo-Huang/academic-research-skills.git ~/.claude/skills/academic-research-skillsOptional — install Python dependencies for faster API access with rate limiting:
pip install requests arxiv huggingface-hubWithout Python, the skill falls back to Claude's built-in WebFetch. Same results, slower throughput.
The skills activate through natural language in Claude Code:
- Paper discovery:
"Find papers on [topic]"— runs the 4-phase discovery pipeline - Full pipeline:
"I want to write a paper on [topic]"— executes all 11 stages from discovery to PDF - Deep research:
"Research the impact of AI on [field]"— 13-agent analysis with gap identification - Peer review:
"Review this paper"— 5-person simulated review with scoring - Research guidance:
"Guide my research on [topic]"— Socratic dialogue with SCR reflection protocol
Phase A: Community Intelligence → trending signals from X, GitHub, HuggingFace
Phase B: Systematic Search → 200-500 candidates via S2 + arXiv + HF + GitHub APIs
Phase C: Verification → each candidate verified against its arXiv page
Phase D: Citation Graph Expansion → forward/backward citation traversal via S2 API
Phase E: Scoring → PIS (Paper Importance Score) ranking
PIS scoring combines citation velocity, venue prestige, community buzz, and recency decay. Weights adapt by paper age — newer papers scored on relevance, mature papers on impact.
DISCOVER → RESEARCH → WRITE → VERIFY → REVIEW → REVISE → PUBLISH
11 stages with two mandatory integrity checkpoints. Every reference verified. Two rounds of peer review.
Complete artifacts from a real pipeline run are available in examples/showcase/:
| Artifact | Description |
|---|---|
| Final Paper | APA 7.0, LaTeX typeset |
| Integrity Report | Caught 15 fabricated references + 3 statistical errors |
| Peer Review | 5-person review panel, 0-100 scoring |
| Post-Publication Audit | Stress test found 21/68 issues after 3 rounds of automated checks |
Model: Claude Opus 4.6 recommended (Max plan). Full pipeline may exceed 200K tokens.
Unattended mode: claude --dangerously-skip-permissions
Optional API keys:
| Key | Purpose | Cost |
|---|---|---|
S2_API_KEY |
Higher Semantic Scholar rate limits | Free |
OPENAI_API_KEY |
Community intelligence via last30days skill | Paid |
academic-research-skills/
├── discovery/ Paper discovery engine (v2.1)
│ ├── scripts/ research_radar.py — Python API client
│ └── agents/
├── deep-research/ 13-agent research team (v2.4)
├── academic-paper/ 12-agent writing pipeline (v2.4)
├── academic-paper-reviewer/ 5-person peer review (v1.4)
├── academic-pipeline/ 11-stage orchestrator (v2.8)
├── shared/ Cross-skill data contracts
└── examples/showcase/ Real pipeline artifacts
See CONTRIBUTING.md.
CC BY-NC 4.0 — Free to share and adapt for non-commercial use with attribution.
Based on academic-research-skills by Cheng-I Wu.