A Claude co-pilot for the entire research lifecycle — from picking a question through final peer review. Designed for academic researchers (graduate students, postdocs, faculty) and applied researchers (UX, policy, public health, data science) who want a rigorous collaborator that respects methodological standards instead of generating plausible-looking output.
Works in two places:
- Claude Code (CLI / IDE) — installs as a plugin with skills, slash commands, and subagents.
- Claude.ai (web / desktop) — install individual skills as Claude Skills.
Same skills, same behavior, two surfaces.
| Stage of research | Skill | Trigger examples |
|---|---|---|
| Finding a question | research-brainstorm | "help me find a thesis topic in X", "brainstorm research ideas" |
| Reviewing the field | literature-review | "what does the research say about X", "do a lit review on Y" |
| Designing the study | methodology-advisor | "should I use an RCT or quasi-experiment?", "what sample size do I need?" |
| Stress-testing the ethics | ethics-committee | "is this study ethical?", "review my IRB application", "draft an ethics statement" |
| Designing instruments | survey-design | "is this question biased?", "find a validated scale for X" |
| Analyzing quantitative data | data-analysis | "clean this dataset", "fit a regression", "run a power analysis" |
| Analyzing qualitative data | qualitative-coding | "code these transcripts", "build a codebook", "find themes" |
| Drafting the manuscript | manuscript-drafter | "write the intro", "draft the discussion", "write the abstract" |
| Replicating someone else's study | replication-designer | "design a replication of X", "is this finding robust?" |
| Funding the work | grant-writer | "draft my specific aims", "NSF broader impacts", "ERC synopsis" |
| Presenting it | talk-builder | "turn this paper into a 12-min conference talk", "job talk outline", "thesis defense slides" |
| Responding to reviewers | reviewer-response | "draft my R1 response", "rebuttal letter", "address reviewer comments" |
| Formatting references | citation-formatter | "format these in APA", "fix my bibliography" |
| Getting feedback | peer-review | "review my paper", "fact-check this draft", "review my slides" |
Every skill is grounded in research methods literature and refuses common AI failure modes — no fabricated citations, no p-hacking, no glossed-over disagreement between sources, no qualitative "themes" without an audit trail, no embellished findings, no voice-mismatched drafts.
Each skill name above links to its per-skill README with a synthetic example, the trigger phrases, composition with other skills, and honest caveats. The examples/ folder contains a worked sample for each shipped skill.
The skills aren't 14 islands — they're an orchestrated network: skills are nodes, shared artifacts are edges, and /research is the conductor. Each skill reads what upstream skills produced, does its one job, writes a predictable artifact, and points at what comes next.
research-brainstorm
└→ literature-review
└→ methodology-advisor ──→ ethics-committee (audit before data)
├→ survey-design (if an instrument is needed)
└→ [DATA COLLECTION — your turn]
├→ data-analysis ──→ stats-validator (independent second look)
└→ qualitative-coding
└→ manuscript-drafter ──→ citation-formatter
├→ peer-review (pre-submission audit)
│ └→ reviewer-response (after an R&R)
└→ talk-builder (conference talk)
grant-writer ← draws on brainstorm + methodology (funding; parallel track)
replication-designer ← draws on a target paper; can seed a fresh cycle
Two ways to use the network:
- Pipeline mode — say
/research start a project on <topic>and the conductor walks the whole lifecycle, invoking each skill in turn (Claude Code), reading and writing the project vault (a clean, templatedresearch/<project>/folder), and pausing at the steps only you can do (collecting data, getting IRB approval, submitting). It resumes where you left off. - Pairwise handoffs — run any single skill; when it needs an upstream artifact or finishes with an obvious next step, it offers to chain to the relevant skill (Claude Code) or tells you which
/commandto run next (claude.ai). It never auto-runs another skill without your go-ahead.
Skills coordinate through the project vault — a clean, visible, templated folder (research/<project>/) where every skill files its output into a numbered lifecycle-stage subfolder (01-ideation/, 02-literature/, … 10-dissemination/), alongside a knowledge/ folder and an auto-generated project README.md. No more hunting for where a file went; the project reads top-to-bottom like its own story. Vault use is opt-in — run any skill standalone and it behaves exactly as before.
research/<project>/
README.md manifest.json knowledge/ 00-inbox/
01-ideation/ 02-literature/ 03-methodology/ 04-ethics/ 05-instruments/
06-data/ 07-analysis/ 08-drafts/ 09-review/ 10-dissemination/
audits/ archive/
Beyond filing the documents, the vault holds the project's knowledge: canonical facts (sample N, IRB#, pre-registration, target journal, language), one shared bibliography every skill cites from, the decisions log, consolidated open questions, the glossary, the voice profile, and entities (pseudonyms only — never PII). Each skill reads the vault at intake (so it never re-asks your sample size or re-fabricates a citation) and updates it at output. /vault audit scans every document against the vault to catch drift before a reviewer does — the abstract that says N=240 while the methods say 247, a citation that doesn't resolve, a term used three ways, PII in the wrong file. /vault organize keeps the folder tidy. Full contract: docs/research-vault.md.
Full network map, manifest schema, the human-gate rule, and the Claude Code ↔ claude.ai degradation table: docs/skill-network.md.
Plain Claude is great at fluency, weak at rigor. Researchers feel this most when:
- It invents DOIs and authors that don't exist.
- It runs statistical tests without checking assumptions.
- It generates "themes" from qualitative data without showing where they came from.
- It writes literature reviews that average away real disagreement in the field.
- It says "more research is needed" instead of identifying specific gaps.
- It drafts manuscript prose that doesn't match the author's existing voice.
Each skill in this plugin is a structured workflow that catches these failures: source verification before citation, assumption diagnostics before reporting effects, codebook + memos for every theme, surfacing of conflicting findings, specific rather than generic gaps, mandatory voice extraction before drafting. The skills are written in the voice of a methodologist who has supervised dissertations, reviewed for journals, and published in mixed methods.
For a longer discussion of the design philosophy, see docs/philosophy.md. For common questions about safety, citation, language coverage, and how this compares to SaaS research-AI products, see docs/faq.md.
You need Claude Code installed and a working terminal.
One-command install (recommended):
# Add this repo as a Claude Code marketplace
claude plugin marketplace add Marazii/research-co-pilot
# Install the plugin from that marketplace
claude plugin install research-co-pilot@research-co-pilot-marketplaceAfter install, run /research for the menu, or invoke any specific skill by name (e.g. /lit-review climate adaptation in coastal cities).
Updating an existing install:
claude plugin marketplace update research-co-pilot-marketplace
claude plugin update research-co-pilot@research-co-pilot-marketplaceThen restart Claude Code.
Local install (for development or to use without GitHub):
git clone https://github.com/Marazii/research-co-pilot.git
cd research-co-pilot
# Add your local clone as a marketplace
# Edit ~/.claude/settings.json and add:
{
"plugins": {
"marketplaces": {
"research-co-pilot-local": {
"type": "local",
"path": "/absolute/path/to/the/parent/of/research-co-pilot"
}
}
}
}
# Then in Claude Code:
claude plugin install research-co-pilot@research-co-pilot-localEach skill is uploaded to claude.ai individually as a Claude Skill. Pre-built upload bundles ship as GitHub Release assets — stable URLs that don't change when commits land on main.
-
Go to the latest release and download the
.zipfor each skill you want.Direct downloads always pointing at the latest stable release:
https://github.com/Marazii/research-co-pilot/releases/latest/download/literature-review.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/methodology-advisor.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/ethics-committee.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/data-analysis.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/qualitative-coding.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/research-brainstorm.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/manuscript-drafter.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/replication-designer.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/grant-writer.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/talk-builder.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/reviewer-response.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/citation-formatter.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/survey-design.zip https://github.com/Marazii/research-co-pilot/releases/latest/download/peer-review.zip -
In claude.ai, open Settings → Capabilities → Skills (Pro/Team/Enterprise plans), or open your workspace's Skills library (Team/Enterprise).
-
Click Upload skill and select a
.zip. Repeat per skill.
Once uploaded, claude.ai loads the skill automatically when your message matches its trigger phrases (e.g. "do a literature review on…" loads literature-review).
Subagents and slash commands are Claude Code-only and don't apply in claude.ai. The skills handle their work directly when run there.
Screenshots, demo GIFs, and the social-preview image live under docs/:
docs/philosophy.md— design choices and what the plugin won't do.docs/faq.md— common questions about citations, regulated data, language coverage, comparison to SaaS products.docs/screenshots/— high-DPI screenshots of typical outputs (annotated PDFs, methodology with the AI/ML extensions table, etc.).docs/demo.gif— ~30-second screencast of a typical flow.
(Some visibility assets are added during the v0.10.0-rc.1 → v0.10.0 final cycle and may not all be present yet.)
/brainstorm I work on misinformation and trust in news, broadly. I have access to
panel survey data and can run online experiments. Constraint: must be defensible
in a quant-focused department.
The skill generates 15-25 candidate questions across descriptive, causal, predictive, and contrarian framings; scores them on interest / answerability / novelty / feasibility; sharpens the top 3 into full study sketches with predicted findings, risks, and follow-up studies. See examples/research-brainstorm/ for a worked sample.
/lit-review the effect of remote work on early-career mentorship in knowledge industries
The skill searches, reads, fact-checks, and synthesizes — organizing by idea (not by source), surfacing where studies disagree, marking each claim with a confidence tag, producing a complete review with appraisal table. See examples/literature-review/ for a worked sample.
/analyze ./data/study2_clean.csv — they want me to handle the clustered structure
(students within classrooms within schools). Currently I just have OLS.
The skill walks through the data, fits the appropriate mixed-effects model, runs diagnostics, compares specifications, and writes up the result with effect sizes, CIs, and an honest discussion of what changes from the original. See examples/data-analysis/ for a worked sample.
/code-themes ./transcripts/
The skill anonymizes, develops a codebook (inductively, deductively, or hybrid — you choose), codes systematically, runs inter-rater reliability if you have a second coder, develops themes with disconfirming-case checks, and writes the qualitative findings section with quoted excerpts and a reflexivity statement. See examples/qualitative-coding/ for a worked sample.
/respond review_round1.md manuscript_v1.md (R1 from J. Hypothetical Studies)
The skill categorizes every reviewer point (concession / partial / pushback / clarification / out-of-scope / minor), drafts the response per point with the reviewer's verbatim quote, drafts the corresponding manuscript revisions (delegating long-form prose to manuscript-drafter to preserve voice), assembles a one-page cover letter that surfaces points of disagreement up front, and self-audits that every claimed revision actually appears in the revised manuscript.
/draft Discussion section, target ~1500 words.
Inputs: existing_draft.md, methodology.md, qual_findings.md, references.bib.
Language: Hebrew.
The skill extracts a voice profile from the existing manuscript (sentence length, person/voice, hedge intensity, signature phrases, connectors, citation style, Hebrew register markers), drafts new prose to match, enforces a per-section word budget, scans for non-academic register, and grounds every new idea in your bibliography (or flags it as [LITERATURE NEEDED] with concrete search guidance). See examples/manuscript-drafter/ for a worked sample showing the v0.8.0 overhaul.
research-co-pilot/
├── .claude-plugin/
│ ├── plugin.json # Claude Code plugin manifest
│ └── marketplace.json # one-command install from GitHub
├── .github/
│ ├── workflows/ # CI: validate / build-dist / release
│ ├── ISSUE_TEMPLATE/ # bug / feature / new-skill-proposal
│ └── PULL_REQUEST_TEMPLATE.md
├── skills/ # Portable: work in Claude Code AND claude.ai
│ ├── literature-review/
│ ├── methodology-advisor/
│ ├── ethics-committee/
│ ├── data-analysis/
│ ├── qualitative-coding/
│ ├── research-brainstorm/
│ ├── manuscript-drafter/
│ ├── replication-designer/
│ ├── grant-writer/
│ ├── talk-builder/
│ ├── reviewer-response/
│ ├── citation-formatter/
│ ├── survey-design/
│ ├── peer-review/
│ └── vault/ # the project's knowledge librarian (facts, bibliography, audit)
│ └── (each skill has SKILL.md + README.md + optional reference files)
├── agents/ # Subagents (Claude Code only)
│ ├── source-finder.md # parallel reading of many academic sources
│ ├── data-cruncher.md # heavy compute in isolation
│ ├── transcript-coder.md # bulk qualitative processing
│ ├── manuscript-drafter.md # long-form drafting in isolation
│ ├── stats-validator.md # independent second-look on a colleague's analysis
│ └── vault-organizer.md # read a pile of files + file them into the vault's categories
├── commands/ # Slash commands (Claude Code only)
│ ├── research.md # /research — entry-point router
│ └── (one per skill: /lit-review, /methodology, /ethics, /analyze,
│ /code-themes, /brainstorm, /draft, /replicate, /grant, /talk,
│ /respond, /cite, /survey, /peer-review, /vault)
├── examples/ # Synthetic minimal example per skill
├── docs/ # skill-network, research-vault, philosophy, faq, screenshots, demo
├── scripts/
│ └── build-zips.sh # rebuild dist/ after editing skills
├── CHANGELOG.md
├── CITATION.cff # academic citation metadata
├── CONTRIBUTING.md
├── CODE_OF_CONDUCT.md
├── SECURITY.md
├── LICENSE # MIT
└── README.md
| Skill | What it does |
|---|---|
literature-review |
Fact-checked synthesis of a body of work. Verifies citations, organizes by idea, surfaces disagreement, marks confidence on every claim, identifies specific gaps. Supports narrative, systematic, scoping, rapid, and thematic reviews. |
methodology-advisor |
Quant + qual research design — picking the right design for the question, sampling strategy, power and sample size, validity threats, IRB/ethics, pre-registration. Includes mandatory creative AI / ML / Big Data extensions section that forces the researcher to consider non-conventional methods. |
ethics-committee |
Simulates an IRB / REC / HREC pre-submission review. Audits informed consent, risk-benefit, vulnerable populations, data privacy, deception, payment, AI/LLM use, social media data, equity in recruitment. Produces a decision letter with required revisions. Optional 3-reviewer panel mode. Not a substitute for institutional approval. |
data-analysis |
End-to-end quantitative work in Python or R: cleaning, EDA, statistical testing, modeling (regression, mixed-effects, predictive, time series, survival), assumption diagnostics, sensitivity analyses, visualization, reproducible scripts. |
qualitative-coding |
Codebook development and application using thematic analysis, grounded theory, IPA, framework analysis, or content analysis. Inter-rater reliability (Cohen's κ, Krippendorff's α). NLP-assisted exploration for large corpora — with required validation against hand-coding. |
research-brainstorm |
Generates 15-25 research ideas via question-form variations, cross-field grafts, and contrarian moves. Scores them. Sharpens the top 3 into study sketches. Pushes past the obvious next study. |
manuscript-drafter |
Drafts long-form manuscript sections from your methodology and analysis outputs. Preserves the existing manuscript's voice via mandatory voice-profile extraction. Enforces hard per-section word budgets with a compression pass. Enforces academic register via a banned-pattern audit. Grounds every new idea in literature (distinguishes [CITATION NEEDED] from [LITERATURE NEEDED]). Two-pass ideation → prose workflow. Drafts in the manuscript's language (English, Hebrew, etc.). |
replication-designer |
Designs direct, conceptual, generalization, or robustness replications. Extracts the original spec, justifies every deviation, computes adequate replication power, plans pre-registration, pre-specifies replication-success criteria, supports multi-site logistics. |
grant-writer |
Drafts proposal sections (Specific Aims, Significance, Innovation, Approach, Broader Impacts, DMP, lay summary, biosketch, budget justification) tuned to NIH (R01/R21/F31/K), NSF (Standard / CAREER / GRFP), ERC, Wellcome, Horizon Europe, or foundation grants. Won't overpromise. Honest about scheme fit. |
talk-builder |
Turns one or more papers into an academic talk — outline, per-slide content, speaker notes, opening hook, single take-home message, backup slides for Q&A, rehearsal plan. Adapts to length, audience, format, and discipline conventions. Deck-platform-agnostic outline plus optional Marp / Quarto / reveal.js / Beamer stubs. |
reviewer-response |
Drafts rigorous, polite, point-by-point response to reviewer comments (R1 / R2 / R3) plus matching manuscript revisions. Categorizes each comment, drafts response + revision, assembles cover letter. Won't concede a point the data doesn't support. Composes naturally with peer-review and manuscript-drafter. |
citation-formatter |
APA 7, MLA 9, Chicago (NB and AD), Harvard, Vancouver, IEEE, AMA, journal-specific. Verifies DOIs, handles edge cases (preprints, datasets, software, AI tools). Generates BibTeX/RIS. Document-wide consistency check. |
survey-design |
Question wording, scale choice, ordering effects, response burden. Recommends validated instruments rather than inventing new ones. Includes a pilot plan (cognitive interviews + quantitative pilot). Translation guidance. |
peer-review |
Multi-mode rigorous review: paper verdict, homework grading, committee panel, fact-check audit, plagiarism check, draft thinking-partner, presentation feedback, post-review iterate. Returns a reviewed file with annotations anchored at the relevant locations across .docx (inline comments + tracked changes), .pdf (sticky-note comments + highlights via PyMuPDF), .pptx (native PowerPoint comments on slides / shapes), .tex (% REVIEWER: line comments, optional changes package). |
vault |
The project's knowledge librarian (infrastructure, not a lifecycle stage). Manages the research vault — canonical facts (sample N, IRB#, pre-reg, journal, language), one shared bibliography, the decisions log, consolidated open questions, glossary, voice profile, and entities (pseudonyms only — never PII). /vault audit scans every document against the vault to catch cross-artifact drift (mismatched sample sizes, unresolved citations, term inconsistency, accidental PII). Files stay the source of truth; the vault is reconciled. |
| Agent | Use it when |
|---|---|
source-finder |
Reading >5 academic sources in parallel without bloating the main conversation. Returns structured digests with verified citations. |
data-cruncher |
Heavy computation, many model variants, simulations, sensitivity grids — in isolation. Returns a tight results summary instead of raw output. |
transcript-coder |
Bulk cleaning, anonymization, and code-application across many transcripts. Returns coded JSON + a summary instead of full transcript text. |
manuscript-drafter |
Long-form drafting (whole sections or whole papers) in isolation so the parent doesn't get flooded with thousands of words. Returns a structured digest plus the draft as files. |
stats-validator |
Independent second-look on a colleague's analysis. Reads their script + data + report in fresh context (no narrative contamination), re-runs, sensitivity-checks, and returns a tight memo with confidence judgment. |
vault-organizer |
Reading through a pile of files (a messy folder, a Downloads dump, an inherited project) and filing them into the vault's categories. Classifies each file → stage folder with confidence + rationale, flags PII (never files real-name keys / signed consent / raw identifiers into the vault), and returns a filing plan; on approval, files non-destructively (copy by default). Spawned by /vault organize <path>. |
/research is the entry point if you're not sure which skill you need — it routes by description, or conducts the whole pipeline. The other commands invoke the matching skill directly: /lit-review, /methodology, /ethics, /analyze, /code-themes, /brainstorm, /draft, /grant, /replicate, /talk, /respond, /cite, /survey, /peer-review, and /vault (project knowledge + drift audit).
- Claude Code: primary deliverable written to a markdown file in your working directory (e.g.
lit_review_<topic>.md,analysis_<topic>.md,methodology_<study>.md). Reproducible scripts saved alongside. - Claude.ai: primary deliverable rendered as a downloadable artifact; code and data files saved in the analysis sandbox for download.
Either way, the deliverables are designed to be portable — markdown + scripts you can drop into Overleaf, Google Docs, R Markdown, Quarto, or your repo.
Each skill's SKILL.md is the single source of truth for its behavior. Edit it to:
- Change trigger phrases (frontmatter
description). - Add discipline-specific guidance (e.g. CONSORT for clinical trials, COREQ for qualitative).
- Adjust the output format to match your lab's or journal's expectations.
After editing skills, rebuild the claude.ai bundles:
./scripts/build-zips.shThen re-upload the changed dist/*.zip files to claude.ai (or push to a tag and CI builds + uploads to the Release automatically).
Contributions are welcome with a high bar. Start with CONTRIBUTING.md — it documents:
- What the project is (and is not).
- The PR-level bar (validation, CHANGELOG, version bump, dist rebuild, per-skill README + example updates).
- The path for new skills (open a
New skill proposalissue first; PRs without prior discussion are declined). - What gets declined and why.
By contributing you agree to the Code of Conduct (Contributor Covenant 2.1). Security issues should follow the SECURITY.md disclosure process.
If you use this plugin in your research, please cite it. The repository's CITATION.cff renders as a "Cite this repository" widget on GitHub, producing APA, BibTeX, and other formats. A Zenodo-minted DOI will be available from the v0.10.0 final release onward.
For a longer write-up of the design choices — why each skill encodes specific hard rules, what "rigor over fluency" means in practice, and what the plugin won't do — see docs/philosophy.md.
- Citations: the literature-review skill verifies what it can, but published research changes. Treat the output as a strong first pass that you read and confirm — never paste it into a manuscript without checking.
- Statistics: the data-analysis skill picks reasonable defaults and checks assumptions, but applied stats often involves judgment calls a model can't make. Treat its output as a draft analysis your statistician reviews.
- Qualitative coding: AI-assisted coding is exploratory. Use it to surface candidates, then validate against hand-coded subsets before relying on labels at scale.
- Peer review: useful for catching blind spots and surface issues. Cannot replace expert review in your specific subfield.
- Ethics and IRB: the methodology and ethics-committee skills flag IRB-relevant considerations but are not a substitute for your institution's review process.
- Voice preservation: the manuscript-drafter and reviewer-response skills extract voice from your existing prose. Without an existing manuscript, voice cannot be enforced.
In short: this is a force multiplier for a researcher who knows what good work looks like — not a replacement for that judgment.
MIT — see LICENSE.
Built by Maya Arazi — all skills (including peer-review), subagents, and slash commands. Methodological frameworks reference standard research methods literature (Braun & Clarke, Charmaz, Shadish-Cook-Campbell, Dillman, Lincoln & Guba, and many others) — all citations belong to their original authors.