Read the whole setupGet the setup that works.
Drop your email and unlock the rest: what you need, the install for a single skill, the update and pin commands, and how to tell if a skill actually fired.
One email. No follow-up unless I ship something that fits.
WHAT YOU NEED- Python 3.13 or newer
- The GitHub CLI (
gh), for the gh skill install path - Claude Code, Cursor, Codex or another Agent Skills-compatible client
- The underlying library, already installed, for whichever skill you pick
INSTALLOne named skill, for your agent
Name the skill and the agent so the install stays small. The full-library install exists but is rarely what you want on your first run.
The install
terminalone skill, one agent
$ gh skill install K-Dense-AI/scientific-agent-skills scanpy --agent claude-code → scanpy skill installed and available in your next session
Repeat the command with a different skill name to add more. Each one stays a separate, removable install rather than one monolithic package.
Prove itAsk the agent a question that needs the skill, for example a scanpy preprocessing question. An answer that cites the skill's documented workflow, not a generic guess, means it loaded.
CONFIGUREPin it if it needs to reproduce
There is no settings file, the two real choices are where you install to and whether you pin a version.
terminalreproducible install
$ gh skill install K-Dense-AI/scientific-agent-skills scanpy --pin v2.66.0
| Setting | Default | Why I change it |
|---|
| Pin | tracks main | A published-analysis workflow should not silently change behavior when the skill updates. Pin the version you validated against. |
| Install scope | user-level (~/.agents/skills) | Project-level (.agents/skills/ inside the repo) when a specific research project needs a skill nobody else on the machine should get by default. |
Updatinggh skill update refreshes one installed skill; gh skill update --all refreshes every skill from this library you have installed.
DAILY DRIVINGAsk in plain words
The skill loads into context automatically once installed. You do not invoke it by name, the agent recognizes when the task matches.
- Ask the domain question. "Run differential expression on this bulk RNA-seq count matrix" is enough for the bulk-rnaseq skill to engage.
- Let it choose the database. Skills that wrap a database, PatentsView, Treasury fiscal data, know the current API shape, including recent breaking changes the maintainers already patched.
- Check the citation. A skill answer sourced from a specific database or paper should say so. If it does not, ask which source it used.
The value shows up as fewer wrong API calls and fewer confidently wrong defaults, not as a new interface to learn.
WHEN IT DRIFTSTwo failures, both configuration
The agent still guesses at the API
The skill was never installed, or was installed for a different agent flag than the one you are running. Re-run gh skill install with the correct --agent value.
A database lookup returns an error
The upstream API changed. Check docs/security-report.md and recent commits, since this project patches dead or changed endpoints quickly, most recently PatentsView and Treasury.
ResetRemove the skill directory under ~/.agents/skills/ or the project-level .agents/skills/ and reinstall. Nothing outside that folder is touched.
THE SHAPE OF ITWhere everything lives
~/.agents/skills/scientific-agent-skills/user-level install, available in every project .agents/skills/scientific-agent-skills/project-level install, scoped to one repo skills/<name>/one folder per skill, inside this repo, each independently testable docs/security-report.mdthe CI security scan results per skill tests/the pytest suite you can run per skill or in full
Removing the top-level folder removes every skill installed from this library in one step.
That is the whole setupOne skill, named and pinned, beats installing all 164 at once. The agent gets the specific documentation it needed for the one library your lab actually runs, and nothing else crowding its context.
Open on GitHub →
↑ Unlocks the moment you submit the form above.
