Multi-Agent Academic Research System for Arts & Social Sciences
Automated research pipeline · Web-based topic scoping · LLM-scored peer review · PDF & Word export
DEMO NOTICE
This is a demonstration project showcasing how Claude Code plugins can automate academic research workflows. The papers it produces are AI-generated drafts for reference only — they have NOT been peer-reviewed by human experts.
Please verify all content, citations, and arguments before using any output in real academic work. AI-generated text may contain hallucinated references, unsupported claims, or factual errors. Always conduct your own research and consult domain experts before submission.
AutoArtsResearch is an automated academic research pipeline built as a Claude Code plugin. It guides users through a structured workflow — from topic selection to a submission-ready manuscript — with a focus on arts and social sciences disciplines.
- Structured Research Pipeline: 9-stage workflow (bootstrap → scoping → corpus → framing → evidence → argument → drafting → review → export) with human approval gates
- Three Research Tracks: Literature review, policy analysis, and comparative case study — cumulative by design (Track C includes A + B)
- Arts & Social Sciences Voice: Writing style tuned for humanities disciplines — interpretive headings, discursive prose, thematic framing rather than STEM-style methodology
- LLM Peer Review: 6-dimension scoring (rigor, evidence, citations, method fit, coherence, contribution) with iterative revision loop
- Multi-Format Export: Generates both PDF and Word (.docx) with academic formatting
- Web Viewer: Browse projects, read papers in the browser, and download exports via a local Flask server
- Human-in-the-Loop: 5 mandatory approval gates ensure the researcher stays in control
- Claude Code CLI installed
- Python 3.10+
- macOS, Linux, or WSL
# Install Claude Code CLI (requires Node.js 18+)
npm install -g @anthropic-ai/claude-codeSee the official installation guide for details and authentication setup.
# 1. Clone the repository
git clone https://github.com/YourUsername/AutoArtsResearch.git
cd AutoArtsResearch
# 2. Create the Python environment
source setup.sh
# 3. Launch Claude Code with the plugin
claude --plugin-dir ./plugin/ar:init # Start a new research project (interactive)
/ar:run workspaces/{PROJECT_ID} # Resume the pipeline from current state
/ar:status # Check all project statuses
/ar:web # Launch the web viewer (http://localhost:5050)
The /ar:init flow:
- Asks for your research topic
- Asks which track to use (literature review / policy analysis / comparative case study)
- Creates a workspace and auto-launches the pipeline
- Runs through scoping → writing → review → export with human gates
/ar:init
|
v
Ask research topic + select track
|
v
Create workspace
|
v
Stage 1: Scoping (WebSearch-based topic refinement)
|
v
[GATE 1] User approves research question & scope
|
v
Stage 2: Corpus Building (15-25 web searches, source audit)
|
v
[GATE 2] User approves source corpus
|
v
Stage 3: Framing (literature synthesis, theory, method card)
|
v
[GATE 3] User approves theory & method
|
v
Stage 4: Evidence (source reading, evidence unit extraction)
|
v
Stage 5: Argument (claim construction, argument tree)
|
v
[GATE 4] User approves argument structure
|
v
Stage 6: Drafting (paper from claims + reference verification)
|
v
Stage 7: Review (6-dimension scoring + revision loop)
|
v
Stage 8: Export (PDF + Word)
|
v
[GATE 5] Done — view in web viewer or download files
| Track | Includes | Best For |
|---|---|---|
| A: Literature Review | Thematic synthesis, gap analysis | Emerging fields, interdisciplinary topics |
| B: Policy Analysis | A + policy document analysis, stakeholder mapping | Government reports, institutional statements |
| C: Comparative Case Study | A + B + cross-case comparison | 2-6 cases, regional analysis |
| Dimension | What It Measures |
|---|---|
| Rigor | Logical soundness, analytical depth |
| Evidence Coverage | Breadth and relevance of sources |
| Citation Quality | Proper attribution, source reliability |
| Methodological Fit | Alignment between method and question |
| Coherence | Structural flow, argument consistency |
| Contribution | Originality, significance to the field |
Pass criteria: overall >= 7.0, rigor >= 8.0, citations >= 8.0.
AutoArtsResearch/
├── .claude/
│ ├── agents/ # Model tier definitions
│ │ ├── ar-heavy.md # Opus — orchestration, debate, review
│ │ ├── ar-standard.md # Opus — scoping, writing, analysis
│ │ └── ar-light.md # Sonnet — retrieval, formatting
│ └── skills/ # 18 agent skills
│ ├── ar-init/ # Interactive project setup
│ ├── ar-orchestrator/ # Pipeline driver (9 stages, 5 gates)
│ ├── ar-scoping/ # Topic scoping & research question
│ ├── ar-retrieval/ # Corpus retrieval (academic, policy, media)
│ ├── ar-source-auditor/ # Source reliability scoring & audit
│ ├── ar-lit-synthesis/ # Literature clustering & theme mapping
│ ├── ar-theory/ # Theoretical framework proposal
│ ├── ar-method/ # Method card generation
│ ├── ar-reader/ # Per-source reading & note extraction
│ ├── ar-evidence/ # Evidence unit construction
│ ├── ar-claim-builder/ # Claim construction & argument tree
│ ├── ar-debate/ # Multi-role structured debate
│ ├── ar-research-writer/ # Academic paper drafting
│ ├── ar-ref-checker/ # Reference verification
│ ├── ar-citation-verifier/ # Citation faithfulness check
│ ├── ar-paper-reviewer/ # 6-dimension quality scoring
│ ├── ar-pdf-exporter/ # PDF + Word export
│ └── ar-progress-monitor/ # Background progress tracking
├── plugin/
│ ├── .claude-plugin/ # Plugin metadata
│ └── commands/ # /ar:init, /ar:run, /ar:status, /ar:web
├── utils/
│ ├── schemas.py # Pydantic models, state machine, ID generation
│ ├── md_to_pdf.py # Markdown → academic PDF (fpdf2)
│ └── md_to_docx.py # Markdown → Word document (python-docx)
├── web/
│ ├── app.py # Flask web viewer (port 5050)
│ ├── templates/ # HTML templates
│ └── static/ # CSS styles
├── workspaces/ # Research project workspaces (gitignored)
├── config.example.yaml # Configuration template
├── setup.sh # Environment setup script
└── CLAUDE.md # System prompt & pipeline documentation
workspaces/{PROJECT_ID}/
├── config.yaml # Project config (topic, track)
├── status.json # Pipeline state tracker
├── sources/
│ ├── academic/ # Academic source records
│ ├── policy/ # Policy source records
│ ├── media/ # Media source records
│ └── audit/ # Source audit reports
├── analysis/
│ ├── scoping/ # Scoping report & summary
│ ├── literature_map/ # Thematic clusters & theory proposals
│ ├── method_cards/ # Method card & summary
│ ├── evidence/ # Evidence units & reading notes
│ ├── claims/ # Claim nodes & argument tree
│ └── debate_logs/ # Debate transcripts
├── drafts/
│ └── research_paper.md # Paper draft
├── reviews/
│ ├── review_summary.md # Review scores & feedback
│ ├── ref_check_report.json # Reference verification
│ └── citation_verify_report.json # Citation faithfulness
├── final/
│ └── research_paper.md # Approved final paper
└── exports/
├── research_paper.pdf # PDF export
└── research_paper.docx # Word export
Launch the built-in web viewer to browse research projects:
/ar:web
Opens at http://localhost:5050 with:
- Project list with status and track badges
- In-browser paper reading (rendered markdown)
- PDF and Word download buttons
- Bibliography Generation: Automated BibTeX/reference list from cited sources
- LaTeX Export: Export to LaTeX format for journal submission
- Automated Scheduling: Cron-based pipeline runs for ongoing research projects
- Vector Store: Semantic search over evidence units and source corpus
MIT License
Built with Claude Code