Claude scientific skills for labs, papers and databases

Claude scientific skills for lab and literature work

Claude scientific skills give an agent domain knowledge and tools for lab and literature work: querying biomedical databases, reviewing papers, reproducing results, writing up findings. The category is small, with ten packs reviewed. Two get an install verdict and eight come with caveats. Most of those caveats concern scale, because several packs hold hundreds of skills.

Domain depth is the main reason to install these. Size is the cost. As the Claude Code docs describe, the agent keeps every installed skill’s description in context, so installing a library of 177 skills to use five of them crowds the list.

What to look for in Claude scientific skills

Check what is real. A skill that wraps a database through code or an MCP server does something a prompt cannot. A skill that is a long prompt about a field only reminds the agent of what it already knows. Our depth score separates these. GitHub stars tell you how popular a repo is. They do not tell you whether a skill calls a real tool or pre-approves Bash.

Check permissions. Science packs often run Python, install packages and call many external services. Of the ten, 1 is clean, 5 are marked “review” and 4 have broad access, usually because skills pre-approve unrestricted Bash or fetch instructions from a server. The security page explains each flag.

Then check the limits the authors state. One pack says outright that its output is not clinical advice, and one writing skill pushes toward a one-sided presentation. Finally, check the license for each skill you plan to reuse.

Recommended scientific skills

• ToolUniverse comes from Harvard’s MIMS lab: a Python SDK, an MCP server and 556 skills that reach thousands of scientific databases and tools. It is Apache-2.0 and states a no-telemetry policy. Status: broad access. The optional hosted path has the agent fetch and follow guides from the vendor’s server, and local use runs an unpinned Python package and edits your MCP client configs.
• Co-Researcher is a 15-skill academic plugin. Its literature review skill checks citations and quotes against OpenAlex, arXiv and Europe PMC with local scripts. See our research category for its longer review.
• Scientific Agent Skills is K-Dense’s library of 177 skills across bioinformatics, chemistry, clinical research, data analysis, lab tooling and scientific writing. It is deep and actively maintained, but it has a caveats verdict. In the skills we read, 23 pre-approve unrestricted Bash. Install only the skills you need.
• RigorPilot Skills uses eleven skills to reproduce deep learning repositories from their README: a reviewed plan, a bounded run and an evidence bundle. It is written defensively, but it executes commands from third-party repositories on your machine without an OS sandbox.
• Anti-Defensive Writing is clean. It is a lightweight Chinese and English skill that restructures paper drafts around the strongest claim. Use it for structure and framing, not as permission to omit real limitations.

Also note Academic Research Skills: very detailed and tested, but licensed CC BY-NC 4.0 per its README, so not for commercial use.

The other packs in this category

Academic Research Skills is five Claude Code skills that cover the academic pipeline: deep research, paper writing, peer review, revision and systematic-review screening. If your work is non-commercial and you want one pack for a full paper cycle, it is the most detailed option here.

ARIS (Auto-Research-In-Sleep) is a large markdown skill set for autonomous machine learning research. It covers idea discovery and experiment runs, and it adds cross-model review loops in which a second model critiques the first. Security status is broad access.

AI Research Skills from Orchestra has 98 skills for AI research and ML engineering, plus an autoresearch orchestrator that runs a two-loop experiment and synthesis cycle. It also has broad access.

ScienceClaw is a research agent system built on the OpenClaw gateway. It ships with 319 skills, and its skills and typed Operators are updated only after a replay-verified execution. You adopt the gateway along with the skills, so it is a bigger commitment than a plain skill pack.

paper-context-resolver is a narrow helper from the RigorPilot repo. When a README leaves a gap, it looks up one reproduction-critical detail in the paper, such as a dataset split.

Comparison of the ten packs

Pack	Focus	Skills	Verdict	Depth	Security
ToolUniverse	Biomedical databases and tools	556	install	4	broad
Co-Researcher	Academic literature review	15	install	4	review
Scientific Agent Skills	Bioinformatics, chemistry, clinical, lab	177	caveats	4	broad
Academic Research Skills	Paper pipeline	5	caveats	5	review
ARIS	Autonomous ML research	large set	caveats	4	broad
AI Research Skills (Orchestra)	AI research and ML engineering	98	caveats	4	broad
ScienceClaw	Agent system on OpenClaw	319	caveats	3	review
RigorPilot Skills	Reproducing deep learning repos	11	caveats	4	review
paper-context-resolver	One detail from a paper	1	caveats	1	review
Anti-Defensive Writing	Paper structure	1	caveats	2	clean

Picking a pack by task

For biomedical lookups against real databases, start with ToolUniverse. Scientific Agent Skills covers more fields, including chemistry and clinical research, and it lets you take individual skills instead of the whole library.

For literature reviews where citations have to hold up, use Co-Researcher, since it checks references against OpenAlex, arXiv and Europe PMC. For drafting, peer review and revision, Academic Research Skills goes further if the license fits your use.

For machine learning experiments, ARIS and Orchestra both automate the research loop. Pick ARIS if you want a second model reviewing the first. Pick Orchestra if you want its experiment and synthesis cycle with a broad ML engineering library behind it. Both have broad access, so run them where unrestricted commands cannot do damage.

For reproducing published results, use RigorPilot, and add paper-context-resolver when the README skips details that only the paper contains.

A sample reproduction workflow

RigorPilot is a good example of how a scientific skill should behave. You point the agent at a deep learning repository. The skill reads the README and writes a plan, which you review before anything runs. The run itself is bounded, and the result is an evidence bundle you can inspect instead of a bare claim that the numbers matched. If the README omits something like the dataset split, paper-context-resolver looks it up in the paper. Because the commands come from someone else’s repository and there is no OS sandbox, use a container or a throwaway machine for this.

Who should not install these

Skip the large libraries if you only need a handful of capabilities. Each installed skill adds a description to context. Skip anything with broad access on a machine that holds credentials or data you cannot afford to expose. Do not rely on any of these packs for clinical decisions. If you are building a commercial product, rule out Academic Research Skills because of its non-commercial license.

For general web research, see research skills. For data work, see data skills. The category page has all ten, and how we review explains the scoring.