A skill is a set of instructions and supporting files for a particular workflow. OpenScience loads relevant skills while working, or you can select one explicitly.
The library has two tiers. The core skills are the research procedures the agent always has on its index: research lookup, literature review, brainstorming, hypotheses, reproduce, autoresearch, compute, delegation, figures, schematics, paper writing (general and ML), citations, peer review, and sources. Their names and one-line summaries are in every Research request, and the agent loads one when the task matches. Everything else, roughly 340 skills covering biology, chemistry, physics, machine learning, data engineering, cloud GPU providers, scientific databases and documents, is the library: found by search, by category, or by exact name when the core index points at it (cloud providers and databases are listed there by name).
Browse the skill directory or your installed version:
openscience skill list --all
Explore the library
The dedicated Skills tab contains the full directory, with a description and usage-instruction link for each bundled entry, plus skill recipes that show the inputs and outputs of common procedures.
Use the directory to discover a procedure, then check the entry available in your installed app. A skill can guide several tools; enabling it does not automatically install every package mentioned by the procedure.
Find and use a skill
Open Customize → Skills. The catalog is organised the way the agent uses it: Core first, in workflow order; Personal for the skills you wrote, installed or keep in the project; the Library as shelves by subject, each folded until you open it, with Activate all / Turn off all per shelf; and Sources, every directory feeding the catalog and any name that lost a collision. The views at the top (All, Core, Library, Personal, Off) narrow the list; search is one flat list across everything. In a conversation, type / for the same tiers, or ask for the procedure in your request.
Use the appropriate single-cell analysis skill to inspect this dataset.
First check the counts and metadata; propose preprocessing before
changing anything.
A skill supplies guidance. It does not guarantee that every referenced package, dataset, or external service is installed or connected.
OpenScience searches the enabled library when it does not know an exact skill name. Search results are suggestions; instructions load only after selecting an available name. If a request includes both an unavailable name and a search query, OpenScience uses the query to find candidates instead of loading a guessed substitute.
The / menu opens on Core, then pinned skills, the session actions, and the whole library by subject; typing filters all of it, and a library skill shows its subject on the right. Selecting a skill prepares its slash command in the composer; it does not send your request. Pin a skill in Customize → Skills to keep it near the top of the menu.
A successful tool load appears in the conversation as Loaded skill: name, even when reasoning and activity are collapsed. Expand it to inspect the saved instructions, then Load details for the source and instruction hash. Searches and failed loads are labeled separately.
Loaded guidance remains available in ordinary follow-up turns through conversation history, so a new load is not required on every turn. The receipt stays on the original turn; it confirms that guidance was supplied, not that every instruction was followed. Context compaction can remove older guidance; load the skill again when its full procedure is needed afterward.
Install from a Git repository
openscience skill add gh:owner/repository
Replace owner/repository with the public repository you intend to install. A full Git URL is also supported. OpenScience reviews discovered skills and reports any that were skipped or rejected. Installation may need a configured model for review.
Installed repositories use a namespace so you can inspect or remove their skills together:
openscience skill list
openscience skill show <namespace>
openscience skill show <namespace>/<skill-name>
Review third-party instructions and any scripts before relying on them. A review is not a guarantee about what an external package or service will do.
Write a skill
openscience skill new leakage-checks --description "Check analysis plans for data leakage" --editor
openscience skill validate leakage-checks
A minimal SKILL.md:
---
name: leakage-checks
description: Check analysis plans for data leakage
category: research
---
Before model fitting:
1. Identify the unit of observation and the train/test split.
2. Check whether preprocessing learned from the test set.
3. Check repeated subjects and near-duplicate records across splits.
4. Report the evidence and unresolved risks before proposing changes.
Use openscience skill edit leakage-checks to revise it. validate --strict also fails on warnings.
Register a local skill directory
A team pack, a vendor's collection or an air-gapped mirror can join the library without editing a config file or restarting. In Customize → Skills, choose Add skill → Add a local folder, enter the path and whether to keep it always (your config), for this project (the project config) or until the server restarts; the folder then appears under Sources with a remove control. The same is available through the API and SDK; the skills are scanned at once, nested category folders included:
curl -X POST http://127.0.0.1:4096/settings/skills/paths \
-H 'content-type: application/json' \
-d '{"path": "/data/team-skills", "persist": "global"}'
persist writes the absolute path to skills.paths in the global or the project openscience.json so it survives a restart; without it the root lives for this project's server process only. A missing or empty directory is rejected, and a directory that is already a root is refused rather than loaded twice. GET /settings/skills/paths lists every root feeding the catalog (bundled, project, user, installed, config, runtime) with the skills it won and the ones it lost to a same-named skill elsewhere; a skill that shadows another carries shadows with the losing paths, so a local edit that had no effect is explained. DELETE /settings/skills/paths?path=… removes a root, POST /settings/skills/reload rescans after files changed outside the app, and GET /skill/{name}/content returns a skill's instructions for clients without filesystem access.
Custom roots load with project precedence: a same-named skill in a registered directory wins over the bundled and personal copies.
Use project skills
Add a skill under .openscience/skills/<name>/SKILL.md, or configure additional folders:
{
"skills": {
"paths": ["./skills", "../shared-skills"]
}
}
Existing .claude/skills/ folders are also discovered. Keep shared procedures in version control so teammates can review changes.
For stable research, record the skill revision alongside the analysis code and inputs. Installing a repository is not the same as permanently pinning its current behavior.
Control the skill picker
openscience skill set-entries <namespace> <entry-one>,<entry-two>
This chooses the entry skills exposed from an installed namespace. The names must match that namespace.
Disable a skill in Customize to hide it from discovery without deleting its files. To uninstall:
openscience skill remove <namespace>/<skill-name>
openscience skill remove <namespace>
See Project instructions for rules that should apply throughout a project instead of only when a skill is used.
Contributors can read the skill runtime design for discovery, source precedence, and the permission checks around loading instructions.
Hugging Face dataset helpers
The bundled hugging-face-datasets skill creates private repositories by default with dataset_manager.py init. Pass --public when you intend to create a public repository. The add_rows and quick_setup commands exit unsuccessfully when validation or row upload fails; check their exit status before continuing an automated workflow.