
# Research Agent Skills
### Turn any AI agent into a research co-scientist
**280+ open-source agent skills that take Claude Code, Codex, Cursor, Gemini CLI, Copilot or any Agent Skills harness from the research question to the accepted paper: literature, hypotheses, experimental design, analysis in AI, ML, biology, chemistry, medicine and physics, writing, citation checks and peer review.**
[](https://skills.sh/kalarislabs/research-agent-skills)
[](#skill-catalog)
[](https://github.com/KalarisLabs/research-agent-skills/stargazers)
[](https://github.com/KalarisLabs/research-agent-skills/commits/main)
[](LICENSE)
[](https://app.devin.ai/org/kalarislabs/wiki/KalarisLabs/research-agent-skills?branch=main)
[](https://github.com/KalarisLabs/research-agent-skills/actions/workflows/validate.yml)
[](https://github.com/KalarisLabs/research-agent-skills/actions/workflows/skill-security.yml)
[](benchmarks/README.md)
[](https://securityscorecards.dev/viewer/?uri=github.com/KalarisLabs/research-agent-skills)
[](https://www.bestpractices.dev/projects/15771)
**Works with**
[](#install)
[](#install)
[](#install)
[](#install)
[](#install)
[](#install)
[](#install)
[](#works-in-any-harness)
```bash
npx skills add KalarisLabs/research-agent-skills
```
One command via skills.sh · no account · pick all skills or only the ones you need · not on npm
**Or paste this into your agent and let it do the install:**
```text
Use the skills in "https://github.com/kalarislabs/research-agent-skills" that are relevant to the current task. Run `npx skills add "https://github.com/kalarislabs/research-agent-skills"` and select the relevant skills, then follow their instructions.
```
Documentation sponsored by Mintlify · code explained by Devin CodeWiki
[Interactive Catalog](https://skills.kalarislabs.com) · [Install](#install) · [Problems solved](#research-problems-this-actually-solves) · [Proof](#proof-a-whole-brain-study-run-end-to-end-with-these-skills) · [Who it's for](#built-for-researchers) · [Benchmarks](#measured-not-claimed) · [Skills](#skill-catalog) · [skills.sh](https://skills.sh/kalarislabs/research-agent-skills) · [Docs](https://docs.kalarislabs.com/research-agent-skills) · [FAQ](#faq)
---
**Research Agent Skills**, maintained by **[Kalaris Labs](https://github.com/KalarisLabs)**, is a free, open-source library of
[agent skills](https://agentskills.io) 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.
Browse and search the entire collection visually at **[skills.kalarislabs.com](https://skills.kalarislabs.com)**.
## Research problems this actually solves
These are the failures researchers hit when they use a general-purpose agent for science, and what the skills do about
each. Numbers come from the committed [benchmark results](#measured-not-claimed); model-based ones are pilots.
| The problem | What the skills do | Evidence |
|---|---|---|
| **Fabricated or wrong citations** | `citation-verification` checks every reference against Crossref, arXiv and DOI-less preprints before it reaches your draft | Flagged 4/4 fabricated and 4/4 corrupted references, passed 10/10 real ones |
| **Text that reads as AI-generated** | `unslop-academic-writing` removes the patterns reviewers recognise | Slop index 83% lower than the same model without it (pilot) |
| **Your agent picks the wrong method** | Each skill states when it applies, so the agent loads the right protocol for the task | Right skill in the top 3 for 96% of 100 held-out prompts, with no false triggers |
| **Journal formatting and submission rules** | Venue skills for Nature, Cell, Science, IEEE, ACM, Springer, Elsevier, APA, PLOS and arXiv | 11 journal-format skills, each with a checklist |
| **Analyses you cannot reproduce** | Skills carry tested scripts, reproducibility statements and explicit verification steps | Every skill passes the static quality gate (mean 95.2/100) |
| **Unsafe third-party skills** | Every skill is scanned for prompt injection, payloads and unreviewed network calls | Cisco skill-scanner, CodeQL and an injection linter run on every pull request |
What it does not do: it will not run your experiments or replace your judgement, and the model-based results are pilots on small samples.
### 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:
| Stage | Your agent can | Skills |
|---|---|---|
| **Question** | Generate and stress-test hypotheses, find gaps | [`hypothesis-generation`](skills/hypothesis-generation/SKILL.md), [`scientific-brainstorming`](skills/scientific-brainstorming/SKILL.md), [`scientific-critical-thinking`](skills/scientific-critical-thinking/SKILL.md) |
| **Literature** | Search papers and databases, run PRISMA reviews | [`literature-review`](skills/literature-review/SKILL.md), [`paper-lookup`](skills/paper-lookup/SKILL.md), [`systematic-review-prisma`](skills/systematic-review-prisma/SKILL.md), [`database-lookup`](skills/database-lookup/SKILL.md) |
| **Design** | Plan experiments, size samples, pre-register | [`experimental-design`](skills/experimental-design/SKILL.md), [`statistical-power`](skills/statistical-power/SKILL.md) |
| **Analysis** | Analyze data with field-standard tools | [`statistical-analysis`](skills/statistical-analysis/SKILL.md), [`scanpy`](skills/scanpy/SKILL.md), [`rdkit`](skills/rdkit/SKILL.md), [`astropy`](skills/astropy/SKILL.md), [`pydicom`](skills/pydicom/SKILL.md), [`pytorch-lightning`](skills/pytorch-lightning/SKILL.md) and 150+ more |
| **Writing** | Draft, visualize and remove AI slop | [`scientific-writing`](skills/scientific-writing/SKILL.md), [`scientific-visualization`](skills/scientific-visualization/SKILL.md), [`unslop-academic-writing`](skills/unslop-academic-writing/SKILL.md) |
| **Verification** | Check every reference and reproducibility claim | [`citation-verification`](skills/citation-verification/SKILL.md), [`reproducibility-statement`](skills/reproducibility-statement/SKILL.md) |
| **Submission** | Format for the venue, prepare arXiv, write grants | [`venue-templates`](skills/venue-templates/SKILL.md), [`arxiv-submission`](skills/arxiv-submission/SKILL.md), [`research-grants`](skills/research-grants/SKILL.md) |
| **Review** | Review papers and answer reviewers | [`peer-review`](skills/peer-review/SKILL.md), [`rebuttal-and-response-to-reviewers`](skills/rebuttal-and-response-to-reviewers/SKILL.md) |
| **Autonomy** | Run supervised, multi-step research loops | [`autoresearch`](skills/autoresearch/SKILL.md) |
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
```mermaid
flowchart LR
A["You ask: