AI Research Skills

Stop asking AI to reread your research project from scratch every
session. Give it skills.

A researcher-facing catalog of 13 verified AI skills that cover the
full research workflow — from finding the first paper to submitting the
final manuscript.

Languages: English | 繁中

What you get: 13 skills covering the full research workflow.
Verification status: 11 of 13 verified end-to-end (T1) against a
real research workspace (1100+ papers in Zotero, live NotebookLM, real
manuscript audits); 2 of 13 verified at the binary / CLI layer (T2);
0 unverified. See docs/verification.md for the
per-skill matrix. Distribution: 9 skills come from one install
(research-hub-pipeline); 4 are standalone clones.

Who this is for: graduate students, PhD researchers, postdocs,
research engineers, librarians, and research support staff who run real
research projects with AI in the loop.

10-Minute First Try

Want to feel what this catalog does before reading the rest?

Prerequisite: Claude Code installed — see https://claude.ai/code.
The skills below activate inside a Claude Code conversation.

Realistic time: ~10 min the first time (Claude Code login +
research-hub install + paper search + matrix build). Subsequent runs
on a configured machine are closer to 2–3 min.

Scenario: "Find 10 papers on my research topic and produce a
comparison matrix I can use to write a literature review."

# 1. Install + onboard (one command — handles persona, skill install,
#    optional NotebookLM login, sample data)
pip install research-hub-pipeline
research-hub setup --persona researcher
#   ↑ pick `analyst` if you don't use Zotero (Obsidian + NotebookLM only)
#     pick `humanities` for qualitative / archival / interpretive work
#     pick `researcher` for the full Zotero + Obsidian + NotebookLM stack

# 2. Find + ingest 10 papers (skip NotebookLM for the first run)
python -m research_hub auto "your topic here" --max-papers 10 --no-nlm

# 3. In Claude Code, ask for the matrix:
#    "Use literature-triage-matrix to compare the 10 papers
#     in cluster <slug-research-hub-just-printed>"

What you get back: .research/literature_matrix.md with 9 columns
— citation, question, method, data, claim, evidence type, limitation,
relevance, where to use the paper. Reproducible reference output:
test-corpus/.../literature_matrix.md
(real 5-paper run on AI agents and social interaction).

Already have research-hub installed but never picked a persona?
Re-run research-hub setup --persona <choice> any time — it's
idempotent.

Pick Your Starting Point

Find the row that matches your immediate goal. Install the named
skills with one command. Skip the rest until you need them.

Helping others adopt AI for research (librarian / RA / advisor)?
No install needed — read docs/install.md and
docs/verification.md and recommend.

Don't see your goal? The full pipeline below covers 8 stages of
research; find your stage and the matching skill.

Full Research Pipeline (click to expand)

The 8 stages of a research project, with the skills that fit each one.
This is the comprehensive reference — most users will pick from the
persona table above and never need to read this.

1. Discover lit  →  2. Organise & compare  →  3a. Frame  →  3b. Plan
        →  4. Build model  →  5. Run & validate (C&V)  →  6. Visualise
        →  7. Draft manuscript  →  8. Submit, respond, wrap up

Three skills don't belong to a specific stage — they're triggered by
task character, not pipeline position. See Cross-cutting tools
below.

1. Discover literature

"What has been done? What should I read?"

Tools you probably use: Zotero · NotebookLM · Obsidian (no native
OneNote skill — use Obsidian as the notes layer.)

2. Organise & compare literature, find the gap

"Where is the research gap? Which 5 papers actually matter?"

3a. Frame the problem

"Is my research question sharp enough to be falsifiable?"

A Socratic dialog partner that asks structured questions to surface
what you'd otherwise leave implicit.

3b. Plan the project (capture the artifacts)

"What am I claiming, with what data, and what's my plan?"

Once 3a has shaped the question, these skills capture the plan as
machine-readable manifests so future AI sessions don't reread the whole
repo.

4. Design and build the model

"What architecture, equations, agents, or prompts do I need?"

Re-read the design_brief.md produced in Stage 3a as your model spec,
then generate implementation scaffolding.

For implementation scaffolding (test harness, plotting, batch edits)
and design review by long-context reading, use the Cross-cutting
tools (codex-delegate, gemini-delegate) below.

5. Run experiments, calibrate, and validate (C&V)

"Is the run reproducible, checkable, extensible? Can I save tokens
across long sessions?"

For repeatable sweeps and post-fix verification, delegate via the
Cross-cutting tools below.

6. Visualise and interpret results

"What does the figure actually show? Does my caption match?"

Tools: matplotlib / plotly / your plotting stack of choice.

7. Draft and revise the manuscript

"Does the prose match the figure? Does it fit the target journal?
Does it sound human?"

For long-form bilingual rewrites or 繁中 / CJK drafts, use the
Cross-cutting tool gemini-delegate below.

8. Submit, respond to reviewers, wrap up

"Are my claims defensible? Is the reviewer response complete? Is the
project state preserved for future me?"

Cross-cutting Tools — Used at Every Stage

Three skills don't belong to a specific stage — they're triggered by
task character:

All 13 Skills

• research-hub — search, ingest, organise papers across Zotero / Obsidian / NotebookLM. (Stages 1, 2)
• literature-triage-matrix — comparison matrix across method, data, claim, limitation. (Stage 2)
• notebooklm-brief-verifier — verify NotebookLM briefs against source bundles. (Stage 2)
• zotero-library-curator — audit Zotero, propose cleanup (preview only). (Stage 2)
• research-design-helper — Socratic dialog through RQ → mechanism → identifiability → validation → risk. (Stages 3a, 4)
• research-context-compressor — .research/ manifests so future AI sessions skip the rescan. (Stages 3b, 5, 8)
• research-project-orienter — fast orientation memo from those manifests. (Stages 3b, 5)
• research-hub-multi-ai — stage-agnostic, character-driven routing across Claude / Codex / Gemini. (Cross-cutting)
• paper-memory-builder — .paper/claims.yml and .paper/figures.yml for manuscript work. (Stage 7)

• academic-writing-skills — manuscript revision, claim-evidence audit, banned-word / humanize, journal format, reviewer response. (Stages 7, 8)
• zotero-skills — full Zotero CRUD, batch metadata, library maintenance. (Stages 1, 2, 7)
• codex-delegate — Claude → Codex CLI handoff for code-heavy work. (Cross-cutting, also Stage 6)
• gemini-delegate — Claude → Gemini CLI handoff for long-context, multilingual, or CJK work. (Cross-cutting, also Stages 6, 7)

Standalone use notes

All 13 skills are usable directly after install — no skill depends
on another skill, and none require a research-hub workspace beyond
what research-hub setup --persona <X> configures for you.

The 1 skill below has a workflow chain worth knowing — not a
dependency, just an order:

• research-project-orienter — reads .research/ manifests for
  speed. If none exist yet, the skill falls back to scanning
  README.md + docs/ (slower); for repeat orientation, run
  research-context-compressor first to produce the manifests.

The other 12 skills work directly with their natural inputs:

• 5 need only Claude Code + your own files: research-design-helper,
  research-hub-multi-ai, research-context-compressor,
  paper-memory-builder, academic-writing-skills.
• 4 need one external service you'd already have: zotero-skills /
  zotero-library-curator (Zotero local API), codex-delegate
  (Codex CLI binary), gemini-delegate (Gemini CLI binary).
• 3 work either with or without research-hub-managed inputs:
  • literature-triage-matrix — paste any paper list (titles + DOIs)
    in chat (per SKILL.md mode #0).
  • notebooklm-brief-verifier — accepts manually-downloaded brief +
    plain source list (per SKILL.md Manual fallback mode, v0.68.2).
    Verified end-to-end
    against a fresh-user setup; produces identical results to the
    research-hub-managed mode.
  • research-hub (knowledge-base) — pick analyst persona for
    Obsidian + NotebookLM only (no Zotero), or humanities for
    Zotero + qualitative defaults.

Install

Prerequisite: Claude Code (https://claude.ai/code).

How skills actually work: each skill is a Markdown instruction
file (SKILL.md) installed under ~/.claude/skills/. AI hosts that
support skills (Claude Code, Cursor with the Claude Code extension,
etc.) automatically read and apply them when your request matches
the skill's trigger description. Skills are not CLI tools or Python
packages — they're prompt scaffolding the host loads on your behalf.

The minimum useful set:

# 1. research-hub — installs 9 skills + onboards your persona in one go
pip install research-hub-pipeline
research-hub setup --persona researcher   # or: analyst | humanities | internal

# 2. academic-writing-skills — for any manuscript work
git clone https://github.com/WenyuChiou/academic-writing-skills \
  ~/.claude/skills/academic-writing-skills

That's it for most researchers. Add as needed:

# Heavy Zotero CRUD (deeper than research-hub bundles)
git clone https://github.com/WenyuChiou/zotero-skills ~/.claude/skills/zotero-skills

# Multi-CLI workflows (Claude + Codex + Gemini)
git clone https://github.com/WenyuChiou/codex-delegate ~/.claude/skills/codex-delegate
git clone https://github.com/WenyuChiou/gemini-delegate-skill ~/.claude/skills/gemini-delegate-skill

# Optional NotebookLM browser automation (handled by `setup` if you
# answer yes when prompted; install separately here if you skipped it)
pip install "research-hub-pipeline[playwright]"
research-hub notebooklm login

Already ran research-hub install --platform claude-code? That
command still works but only writes the SKILL.md files — setup is
the recommended onboarding because it also handles persona, Zotero
default collection, and NotebookLM login. Both are idempotent; you
can run setup any time.

Don't want research-hub at all?

Skip pip install research-hub-pipeline and clone only the standalone
repos you need. You'll get up to 4 skills, fully usable, with no
platform install:

# Manuscript revision + reviewer response (most-used)
git clone https://github.com/WenyuChiou/academic-writing-skills ~/.claude/skills/academic-writing-skills

# Deep Zotero CRUD
git clone https://github.com/WenyuChiou/zotero-skills ~/.claude/skills/zotero-skills

# Hand code-heavy tasks to Codex CLI
git clone https://github.com/WenyuChiou/codex-delegate ~/.claude/skills/codex-delegate

# Hand long-context / 繁中 work to Gemini CLI
git clone https://github.com/WenyuChiou/gemini-delegate-skill ~/.claude/skills/gemini-delegate-skill

The other 9 skills (literature-triage, design-helper, compressor,
orienter, paper-memory-builder, etc.) ship together inside
research-hub-pipeline and aren't separately installable today.

Full guide: docs/install.md. Upgrading from
research-hub-pipeline ≤ 0.45? See the upgrade note in that file.

Verified

All 13 skills exercised end-to-end on a real research workspace
(11 T1 functional + 2 T2 binary; 0 failures). Per-skill matrix +
test-corpus artifacts: docs/verification.md.

Status & License

Lightweight catalog. Each skill is maintained in its canonical repo —
this catalog is the index, not a monorepo.

License: MIT. Contributions welcome — open an issue or PR. New skill
proposals should target either research-hub (workflow integration)
or a standalone repo (deep, single-purpose CRUD).