Extension Path: For Researchers

🚀 Computational researchers (can run Python scripts, have an API key, and can use git) can jump into the advanced path directly. Non-programming researchers (humanities/social sciences, clinical research, literature-first work) can start with literature Q&A (NotebookLM) and Zotero AI tools, then read resources/setup-guide.en.md A-C when needed.

← Back to main path README · Continue here after Track A's A3 or Track B's Stage 7. Apply agentic AI to research workflows.

#Use Cases

Research days break into stages, and AI plays a different role at each stage. Use this table to orient yourself:

💡 Computational vs non-programming researchers: the recommended tools run from light to heavy. Non-programming researchers can usually stop at the first tool in each row; computational researchers should move right only when they need automation.

#Curated Projects

💡 Want to wire Claude Code into NotebookLM, Obsidian, Notion, Excel, PDF, Excalidraw, and other research tools? 81+ integrations in resources/mcp-skills-catalog.en.md (grouped by use case). The section below keeps research-specific tools and marketplaces.

#Research Workflow Marketplaces

flonat/claude-research ⭐⭐⭐

Claude Code infrastructure for PhD researchers — skills, agents, hooks, rules for academic workflows. Strong LaTeX/bibliography focus.

#Literature RAG / Q&A

Future-House/paper-qa ⭐⭐⭐⭐⭐

What it teaches: PDF Q&A designed for citation-grounded Q&A — every answer includes sentence-level citations to reduce hallucination risk. Actual accuracy depends on document type; use the official benchmarks / papers as the reference.

Best for: Researchers writing literature reviews who need "every answer must be traceable to its source." More rigorous than generic RAG.

assafelovic/gpt-researcher ⭐⭐⭐⭐

What it teaches: Autonomous deep-research agent — planner + multi-source crawl + report synthesis. Give it a research topic, get a markdown / PDF brief out.

Best for: Researchers who need to quickly scope new topics and produce research briefs.

#Outline & Writing

stanford-oval/storm ⭐⭐⭐⭐

What it teaches: Multi-perspective outline-then-write pipeline — plain-language version: (1) simulate different perspectives asking questions, (2) organize those questions into an outline, then (3) generate a Wikipedia-style draft. From Stanford OVAL.

Best for: Learning outline-driven writing. Great for producing topic briefs from scratch; the closest open-source analog to NotebookLM's structured report flow.

Notes: Last push was over 6 months ago — verify the latest commit date before relying on it.

kaixindelele/ChatPaper ⭐⭐⭐⭐⭐ (Chinese readers)

What it teaches: Full arXiv workflow for Chinese researchers — paper summary + translation + polishing + review-response generation. Maintained by a Chinese team; defaults are friendly to Chinese-language workflows.

Best for: Chinese graduate students looking for a Chinese-friendly entry-level paper workflow tool.

Notes: License is custom non-commercial — read the original terms before any use; common practice is research / personal use, but you should verify the terms yourself.

#Citation Manager Integrations

MuiseDestiny/zotero-gpt ⭐⭐⭐⭐

What it teaches: A Zotero LLM plugin — chat with your library, summarize selections, generate inline notes.

Best for: Heavy Zotero users who want AI inside their reading workflow without switching tools.

Notes: AGPL-3.0 license (copyleft) — derivative products that ship modifications must follow the terms.

#Multi-LLM Research Stack (Maintainer Setup)

Some research tasks only need Claude (dialogue, design, review). Others waste Claude tokens (large code refactors, long-form drafts). The maintainer's actual setup is Claude as planner / reviewer, Codex for code, and Gemini for long drafts. Use this table to decide which model to use when:

Maintainer's 6 self-used research skills

⚠️ Disclosure: The following 6 tools are research skills used day to day by the maintainer @WenyuChiou (Lehigh CEE PhD candidate) and published for people with similar needs. They have not been independently evaluated by third parties. Best fit: PhD dissertation writing and cross-paper literature organization. They may not fit your field. Full entries are in resources/mcp-skills-catalog.en.md 13 + 14.

#Multi-Agent for Research

langchain-ai/open_deep_research ⭐⭐⭐⭐⭐

What it teaches: Open-source Deep Research — supports both single-agent and supervisor + multi-researcher architectures (the multi-agent path currently lives in src/legacy/), parallel search, citation-grounded report synthesis. A solid reference for "LLM agent that auto-produces a cited brief."

Best for: Researchers building "agent auto-generates a cited brief" workflows. A solid open-source pick when you want a maintained reference implementation.

Notes: Depends on LangGraph + search tools (API key required).

SakanaAI/AI-Scientist-v2 ⭐⭐⭐⭐

What it teaches: End-to-end multi-agent science loop: ideate → code → experiment → write → peer-review. Sakana AI's research implementation of "AI writes a full ML paper."

Best for: Researchers who want to see "what does a swarm of agents running a full research lifecycle look like." Architecture reference, not a production tool.

Notes: Outputs are demo-level (not field-ready), ML/CS-domain bias. License is a custom source-available term (with a manuscript-disclosure clause) — read the LICENSE file before use.

Still missing: actively-maintained peer-review automation, conference-review pipelines. If you've built or know of one, please open a PR.

#Required Reading

1. The Effortless Academic — Claude Code beginner guides
2. Pedro Sant'Anna — Researcher setup guide

#Workflows to Master

The biggest mistake researchers make with AI is opening ChatGPT only when they get stuck. The key is making AI a daily tool by setting a cadence. The 7 workflows below are ordered by usage frequency and are routines the maintainer actually runs, not hypotheticals.

💡 Starter playbook: run the daily inbox triage and writing sprint for one month first. Add advanced workflows only after the habit sticks.

#Tier Recommendations

Researchers do not need to install Claude Code on day one. This is the recommended progression:

Most researchers can stop at Tier 1-2. Tier 3 is worth it only when you have a lot of repeated workflows, such as running the same paper synthesis every week.