A community collection of neuroimaging skills for LLM coding agents.
skills-comm gathers Agent Skills
that teach LLM agents (Claude Code and other agents that read SKILL.md/AGENTS.md)
how to drive real neuroimaging tools correctly and reproducibly. Instead of letting
an agent improvise commands for FSL, ANTs, FreeSurfer, DataLad, nipoppy, or a SLURM
cluster, a skill encodes the operational knowledge — the right flags, the safe order of
operations, the quality-control checks, and the reproducibility guardrails — so the work
is repeatable across people and machines.
The goal is operationalization: turn "how an expert actually runs this tool" into a package an agent can load on demand.
- Method developers & tool authors who want their software used the way it was meant to be used, even when an LLM is at the keyboard.
- Neuroimaging researchers who want a coding agent that follows community best-practice (BIDS, YODA/DataLad provenance, explicit tool versioning) by default.
- Contributors who want to share a reproducible recipe as a reusable skill.
Each contribution lives under plugins/ as its own directory. A contribution
is either a full Claude Code plugin (with a .claude-plugin/plugin.json manifest) or a
bare skill collection (one or more SKILL.md directories). Current contributions:
| Plugin | What it operationalizes |
|---|---|
plugins/datalad-cli |
DataLad / git-annex provenance and YODA-compliant analysis datasets |
plugins/nipoppy-cli |
The nipoppy dataset-management lifecycle (curation → BIDS → process → extract) |
plugins/brain-extraction |
Brain extraction (skull stripping) and its quality control |
plugins/combat-harmonization |
ComBat-family harmonization strategy, execution, and QC |
plugins/fc-matrix-generation |
Functional connectivity matrix generation from resting-state fMRI |
plugins/job-monitor |
SLURM/PBS/LSF job monitoring and fast failure detection |
plugins/registration |
Image registration and registration QC |
skill-iterations/ is a teaching example: the same
brain-extraction skill captured at six successive stages of refinement, showing how a raw
draft is hardened into a reliable skill (fixing YAML, removing "Claude-isms", generalizing
across tools, improving QC). Read it to see what "good" looks like before you write your own.
A skill is a directory containing a SKILL.md with YAML frontmatter and optional
references/ files that the agent loads only when needed:
my-skill/
├── SKILL.md # frontmatter (name, description) + the procedure
└── references/ # detailed, load-on-demand material (tool flags, QC criteria, ...)
The description field is the trigger: it should list the phrases and intents that
should activate the skill. The body is a terse, imperative procedure — not a tutorial.
Install a plugin into Claude Code (session-only for testing, or permanently):
# Try it for one session
claude --plugin-dir ./plugins/datalad-cli
# Install permanently
claude plugin install ./plugins/datalad-cliSee each plugin's own README.md for its skills and slash commands.
We welcome new skills. Open an issue first to propose the skill, then send a PR. Read CONTRIBUTING.md for the full procedure (issue → structure → manual test with evidence → review). Agents contributing on a human's behalf should also read AGENTS_CONTRIBUTING.md and AGENTS_GIT.md.
Apache License 2.0 — see LICENSE. By contributing you agree your contribution is licensed under the same terms.