kalarislabs/research-agent-skills

---

Research Agent Skills, maintained by Kalaris Labs, is a free, open-source library of agent skills that gives your AI agent the methods of a careful scientist. Each skill is a folder with a SKILL.md playbook, reference material and tested scripts. Your agent loads one only when the task needs it, so the same agent can search the literature in the morning, design an experiment at noon and answer reviewers at night.

From assistant to co-scientist

An assistant drafts what you ask for. A co-scientist works through the research itself and checks its own output. The skills cover every stage:

It is built against the three ways AI research help goes wrong:

1. AI slop. unslop-academic-writing finds stock phrasing (“delves into”, “pivotal role”), empty emphasis, stacked hedges and monotone rhythm, then rewrites toward specific, evidence-backed prose in the author’s own voice.
2. Hallucinated citations. citation-verification checks every reference against Crossref, OpenAlex and arXiv. On our labeled benchmark it flagged every fabricated and corrupted reference and passed every correct one.
3. Outdated venue rules. Journal skills make the agent read the current author guidelines before formatting, instead of trusting memory.

How it works

flowchart LR
    A["You ask:'Check every reference in refs.bib'"] --> B["Your agent(Claude Code, Codex, Cursor…)"]
    B -- "reads short skill descriptions" --> C{"Which skill fits?"}
    C -- "loads only that one" --> D["citation-verificationSKILL.md + scripts"]
    D -- "queries" --> E[("Crossref · OpenAlex · arXiv")]
    E --> F["Report: verified,mismatched, not found"]

Installed skills cost almost nothing until used: the agent sees one short description per skill and opens the full playbook only for the task at hand.

Proof: a whole-brain study, run end to end with these skills

Reviving a model fly brain is a Kalaris Labs preprint on anoxia, brain death and resuscitation in a connectome-constrained simulation of the whole adult Drosophila brain (139,255 neurons). Claude worked through Research Agent Skills and Kalaris Labs Myriad from the first hypothesis to the finished paper, under the author’s direction (disclosed in the paper; Claude is not an author).

Scope: all results are predictions of a simplified simulation (connectome-constrained LIF neurons and phenomenological metabolism, constants not fitted to Drosophila). They are not evidence about living tissue, clinical brain death or human resuscitation, and resuscitation of a brain-dead organism has not been demonstrated. Claims and numbers follow the study’s KEY_CLAIMS.json; see also its limitations.

Install

One command with the open skills.sh CLI. It needs Node.js 18+ and no account, and works with the agents above plus about 70 others (Cline, Goose, Roo, Zed, Warp and more).

npx skills add KalarisLabs/research-agent-skills

That opens a picker for skills and agents. To skip the picker:

Then ask your agent in plain language, for example “Verify every reference in refs.bib” or “Unslop this introduction without changing any claims”. The agent loads the right skill on its own.

Choose a starting skill for your field:

Field bundles in Claude Code. The repository is also a Claude Code plugin marketplace with curated bundles:

/plugin marketplace add KalarisLabs/research-agent-skills
/plugin install biology-research@research-agent-skills

Bundles: research-essentials, ml-research, ai-research, biology-research, chemistry-research, medicine-research, physics-research, plus one plugin per category (for example literature-review). Browse all categories when you need a specialist skill.

Works in any harness

The skills follow the open Agent Skills specification: plain folders with a SKILL.md file, no SDK, no server, no lock-in. Every skill passes the specification’s reference validator.

Tool permissions and advanced metadata may vary by harness.

Built for researchers

Find an agent skill for your research task

Explore scientific agent skills by discipline

For more workflows and installation choices, see skills by field.

What you can ask

"Draft the Results section from results/ and the figures in figs/, then unslop it."
"Verify every reference in refs.bib and tell me which ones don't exist."
"Reformat this manuscript for Cell: Summary, Highlights, eTOC blurb and STAR Methods."
"Run a PRISMA 2020 systematic review on exercise and depression: search strings, dedup, flow diagram."
"Write a point-by-point response to these three reviewers within the rebuttal limit."
"Build a citation graph of my Zotero library and find the papers that bridge the two clusters."
"Check my LaTeX source before I upload it to arXiv."
"Analyze this scRNA-seq dataset with scanpy and write the methods section."

Measured, not claimed

Every skill is scored on a static quality rubric and a trigger-routing suite, and task outcomes are measured for the original skills. Every chart below is rendered from the committed results in benchmarks/results/ by benchmarks/charts.py, and CI fails if a chart drifts from its data. Methodology and caveats: benchmarks/README.md.

Model-based results are pilots (small samples, one small model as answerer and judge). Quality and routing gates run on every pull request, so a weak or overlapping skill can’t be merged silently.

Featured skills

Journal and venue formats

Nature · Nature Communications · Scientific Reports · Science · Science Advances · Cell · Neuron · Cell Reports · IEEE Transactions · IEEE Access · ACM (sigconf, acmsmall: CHI, KDD, SIGGRAPH) · Elsevier (elsarticle, CAS) · Springer LNCS · PLOS ONE · PLOS Biology · APA 7th edition · arXiv · NeurIPS · ICML · ICLR · ACL/ARR · AAAI · COLM · OSDI/NSDI/ASPLOS/SOSP · NIH and NSF grants · conference posters and slides.

Skill catalog

Grouped by category and generated from catalog/skills.json. The documentation directory has a page per skill.

Documentation

• Getting started: Installation · Quickstart
• Guides: Write a research paper with AI · Thesis and dissertation · No-slop academic writing · Systematic reviews · Skills by field
• Reference: CLI · Architecture · Benchmarks · Security · Contributing
• For AI search and agents: llms.txt · llms-full.txt (every skill with tags) · catalog/skills.json

Find us on

Security

Skills are instructions and code that your agent runs with your permissions, so they are treated like software dependencies. Every pull request runs spec validation, a prompt-injection and payload linter over all files, the Cisco AI Defense skill-scanner, CodeQL, Semgrep, Bandit, secret scanning and dependency review. Scripts may only contact reviewed domains. Releases are checksummed, signed and attested, and the installer detects tampered files. Details and vulnerability reporting: SECURITY.md.

FAQ

More in the FAQ.

Contributing

Researchers are welcome to contribute skills for their field. Start with CONTRIBUTING.md and the research-skill-creator skill. make check runs everything CI runs. See also GOVERNANCE.md and SUPPORT.md.

Citation

If Research Agent Skills helped your work, please cite it (“Cite this repository” on GitHub uses CITATION.cff):

@software{chowdhury_research_agent_skills_2026,
  author  = {Chowdhury, Sayan and {Kalaris Labs}},
  title   = {Research Agent Skills: AI agent skills for academic writing and scientific research},
  year    = {2026},
  version = {1.1.1},
  url     = {https://github.com/KalarisLabs/research-agent-skills},
  license = {MIT}
}

License

MIT © 2026 Kalaris Labs. Portions of some skills are adapted from third-party MIT-licensed projects, and their notices are in THIRD_PARTY_NOTICES.md.

---