ai-hub-skill-set

AIHubSkillSet

A Claude Code plugin marketplace bundling three composable skills for medical-imaging dataset work: dataset acquisition, DICOM → NIfTI conversion, and nnUNet v2 formatting.

The three skills are independent — Claude loads each one on demand based on the task — but they are designed to compose: typically you acquire data, then convert DICOM to NIfTI, then format for nnUNet training.

Install (recommended — auto-update on /plugin update)

/plugin marketplace add Bardli/AIHubSkillSet
/plugin install ai-hub-skill-set@ai-hub-skill-set

Then /plugin list to confirm and /plugin update ai-hub-skill-set@ai-hub-skill-set whenever you want to pull new revisions.

The three skills

Skill	Purpose	Triggers on
dataset-acquisition	Download from TCGA/GDC, Kaggle, HuggingFace, Google Drive; generate SLURM sbatch scripts	"grab the CESC slides", "pull this Kaggle competition", "download from HF and pin the revision", "sbatch script for this download"
dicom-converter	DICOM series → NIfTI; RTSTRUCT/SEG handling; SOP-UID-anchored routing for multi-acquisition data; 10-check audit script; multi-RTSTRUCT OR-union; debug recipes for label misalignment	"convert these DICOMs to NIfTI", "handle this RTSTRUCT", "debug this label/image mismatch", "audit this DICOM dataset"
nnunet-converter	Format imaging datasets into nnUNet v2 layout (imagesTr/labelsTr/dataset.json/splits_final.json); handles 2D PNG/BMP/TIFF, 3D NIfTI/MHA/NRRD, 3D TIFF, multi-modal, classification labels, ignore label, region-based	"make this nnUNet-ready", "prepare for nnUNet training", "generate dataset.json"

Typical pipeline

        ┌──────────────────────┐
        │ dataset-acquisition  │   download from TCGA/GDC, Kaggle, HF, gdrive
        └─────────┬────────────┘
                  ▼
        ┌──────────────────────┐
        │ dicom-converter      │   DICOM → NIfTI (only if input is DICOM)
        └─────────┬────────────┘
                  ▼
        ┌──────────────────────┐
        │ nnunet-converter     │   nnUNet v2 layout + dataset.json + splits
        └──────────────────────┘

Skip any step that does not apply. If the source data is already NIfTI/MHA, skip dicom-converter. If you do not need nnUNet, skip nnunet-converter.

Manual install (without the marketplace)

If you prefer to vendor the skills directly into a project (no auto-update):

git clone https://github.com/Bardli/AIHubSkillSet.git
mkdir -p ~/.claude/skills            # or .claude/skills/ for project-scoped
cp -r AIHubSkillSet/skills/* ~/.claude/skills/

You can also copy individual skills if you only want one or two:

cp -r AIHubSkillSet/skills/nnunet-converter ~/.claude/skills/

Repository layout

AIHubSkillSet/
├── .claude-plugin/
│   ├── marketplace.json          # marketplace declaration
│   └── plugin.json               # this repo IS the plugin (one plugin, three skills)
├── skills/
│   ├── nnunet-converter/         # progressive-disclosure skill
│   │   ├── SKILL.md
│   │   ├── references/           # 10 topical .md files
│   │   └── scripts/              # convert_template, simple-CLI, manifest writer
│   ├── dataset-acquisition/      # progressive-disclosure skill
│   │   ├── SKILL.md
│   │   ├── references/           # tcga_gdc, kaggle, huggingface, google_drive, sbatch_template
│   │   └── scripts/              # gdc_manifest.py, hf_download.py
│   └── dicom-converter/          # progressive-disclosure skill
│       ├── SKILL.md
│       ├── references/           # 9 topical .md files (audit, SOP-UID routing, multi-RTSTRUCT, etc.)
│       └── scripts/              # audit_dicom_dataset, build_sop_to_acq, parse_rtstruct_union
└── README.md

Design notes

• Progressive disclosure. All three skills use a compact SKILL.md entry point that loads detailed references/*.md on demand, with mandatory MUST read pointers for the references the model has to consult before writing code or commands. This keeps the always-loaded context small while preserving the depth of each topic.
• Strict scoping between skills. Each skill stays in its lane and points at the others when a request crosses boundaries (e.g. dataset-acquisition refuses to do DICOM→NIfTI; it tells you to use dicom-converter).
• Bundled scripts have attribution. Where scripts are adapted from ryanwangk/medimg_skills (MIT), the script header records the source.

Related repositories