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.mdA-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:
| Stage | Common pain point | How AI helps | Recommended tools (light to heavy) |
|---|---|---|---|
| Literature exploration | You do not know the classic papers in a field | Recommendations + summaries + comparison | NotebookLM → paper-qa → gpt-researcher |
| Close reading | You lose the thread halfway through a PDF / miss the claim | Extract claims, figures, citations, and notes | Zotero + zotero-gpt → zotero-skills |
| Research design | The RQ is fuzzy, or the method choice is unclear | Clarifying dialogue and trade-off mapping | Claude.ai chat → ai-research-skills |
| Experiments / coding | Boilerplate repeats and plotting eats time | Write / edit code and batch refactor | Claude Code → codex-delegate |
| Manuscript writing | Drafts stall or sentences do not land | Outline → paragraphs → polishing | Claude.ai → gemini-delegate (long drafts) |
| Revision / submission | Journal requirements are easy to miss | banned-word / figure-text / submission checklist | academic-writing-skills |
| Cross-paper synthesis | Five papers need to talk to each other and context explodes | Read 1M tokens at once and organize the synthesis | gemini-delegate |
💡 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? 62 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 ⭐⭐⭐⭐⭐
| Field | Value |
|---|---|
| Stars | ★ 8k+ |
| License | Apache-2.0 |
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 ⭐⭐⭐⭐
| Field | Value |
|---|---|
| Stars | ★ 27k+ |
| License | Apache-2.0 |
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 ⭐⭐⭐⭐
| Field | Value |
|---|---|
| Stars | ★ 28k+ |
| License | MIT |
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)
| Field | Value |
|---|---|
| Language | Chinese + Python |
| Stars | ★ 19k+ |
| License | NOASSERTION (custom non-commercial) |
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 ⭐⭐⭐⭐
| Field | Value |
|---|---|
| Stars | ★ 7k+ |
| License | AGPL-3.0 |
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:
| Task type | Example | LLM to use | Why |
|---|---|---|---|
| Research design / hypothesis discussion | "Should this RQ use logistic vs survival?" | Claude.ai chat | Collaborative dialogue and context memory |
| Writing / editing code | "Add logging to 50 simulation scripts" | codex-delegate | Fast mechanical edits without burning Claude tokens |
| Long-form drafting (Chinese / English) | "Draft an 8-page paper section" | gemini-delegate | 1M context and strong long-form prose |
| Second opinion | "Ask Gemini to review my discussion section" | gemini-delegate | LLM-vs-LLM comparison makes Claude's own biases easier to spot |
| Pre-submission audit | "Run banned-word + figure-text checklist" | academic-writing-skills | Structured audit instead of ad hoc LLM judgment |
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.md13 + 14.
| Tool | Best for stage | One-liner |
|---|---|---|
| ai-research-skills ⭐⭐⭐⭐⭐ | Full pipeline | 14 research skills packaged as a 5-plugin marketplace; one command installs the set |
| research-hub ⭐⭐⭐⭐ | Literature organization | Zotero + Obsidian + NotebookLM workspace with CLI / MCP / REST / dashboard interfaces |
| zotero-skills ⭐⭐⭐⭐ | Reference management | Zotero CLI skill for search / add / classify / tag; complements zotero-gpt, which chats inside Zotero while this operates from outside |
| academic-writing-skills ⭐⭐⭐ | Pre-submission | banned-word audit, figure-text coupling, and submission checklist; per-paper journal_format / style_overrides customization |
| codex-delegate ⭐⭐⭐⭐⭐ | Coding | Standard Claude planner + Codex executor skill for batch refactor / boilerplate / migration work |
| gemini-delegate-skill ⭐⭐⭐⭐ | Long drafts / synthesis | Claude planner + Gemini for 1M-context long-form writing / CJK / second opinions |
Multi-Agent for Research
langchain-ai/open_deep_research ⭐⭐⭐⭐⭐
| Field | Value |
|---|---|
| Stars | ★ 11k+ |
| License | MIT |
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 ⭐⭐⭐⭐
| Field | Value |
|---|---|
| Stars | ★ 6k+ |
| License | The AI Scientist Source Code License (source-available, non-commercial + manuscript-disclosure clause) |
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
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.
| Frequency | Workflow | How to run it (≤ 3 steps) | Recommended tools | Best for |
|---|---|---|---|---|
| Daily | Literature inbox triage | (1) Put yesterday's papers into paper-qa (2) Extract claims + a 4-5 line summary (3) Move notes into Zotero / Obsidian | paper-qa + zotero-gpt | All researchers |
| Daily | Writing sprint (25 min) | (1) Give one paragraph to Claude.ai (2) Run banned-word + figure-text audit (3) Merge the revision into the main draft | Claude.ai + academic-writing-skills | Paper-writing stage |
| Weekly | Cross-paper synthesis | (1) Feed 5-10 PDFs to Gemini (2) Ask where the papers disagree (3) Turn the answer into a 1-page brief | gemini-delegate (1M context) | Computational researchers |
| Weekly | Zotero cleanup | (1) Mark unread / read (2) Retag items (3) Pull out PDFs that should be archived | zotero-skills or zotero-gpt | All researchers |
| Monthly | Research progress brief | (1) Pull recent notes from Obsidian + Zotero + NotebookLM (2) Summarize 5 progress points (3) Send to your advisor | research-hub | People using all 3 tools |
| Per paper | Final pre-submission audit | (1) banned-word audit (2) figure-text coupling check (3) submission checklist | academic-writing-skills | Final week before submission |
| Per paper | Multi-agent peer review | (1) Claude reviews logic / argument (2) Codex checks code / table numbers (3) Gemini reviews prose / clarity | codex-delegate + gemini-delegate | Pre-submission second opinion |
💡 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:
| Tier | Tools | Best for | Learning cost |
|---|---|---|---|
| Tier 0 | Claude.ai web + NotebookLM | Non-programming researchers, humanities / social sciences, clinical research | 0 (browser skills are enough) |
| Tier 1 | Claude Desktop + Zotero MCP / Obsidian MCP | Researchers already using Zotero / Obsidian | Half-day setup |
| Tier 2 | Claude Code + ai-research-skills | Computational researchers who mostly write / edit code | 1-2 days to get started |
| Tier 3 | Claude Code + codex-delegate + gemini-delegate + research-hub | People building a multi-LLM research pipeline across multiple tools | 1 week setup + ongoing tuning |
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.