Two skills for asking natural-language questions over real public-domain healthcare data.
trove ships two Claude Code skills packaged as a single plugin. Each skill teaches Claude how to query a specific healthcare dataset, how to interpret the fields, and how to cite the source documents — so answers are grounded in real data rather than what an LLM half-remembers from training.
Add the trove marketplace, then install the plugin:
/plugin marketplace add cbetz/trove
/plugin install trove@troveOr copy the skill directories directly if you'd rather skip the plugin layer:
git clone https://github.com/cbetz/trove
cp -r trove/skills/fda-analyst ~/.claude/skills/
cp -r trove/skills/hcris-analyst ~/.claude/skills/Both work in Claude Code, Claude Desktop, and any other tool that loads the Claude Code skill format.
fda-analyst
Answer questions about FDA novel drug approvals (2021–2024) — the basis for approval, the trials, the endpoints, the regulatory pathway, adverse events the FDA flagged in the medical review. The skill fetches the actual approval-package PDFs at query time rather than relying on training data.
hcris-analyst
Answer questions about U.S. nonprofit hospital financials, beds, staffing, charity care, and community benefit by querying CMS Medicare Cost Reports (HCRIS) and IRS Form 990 Schedule H side-by-side. Handles cross-form alignment, definitional caveats, and home-county Social Vulnerability Index context.
A Claude Code skill is a folder of instructions and reference documents that teaches Claude how to do a specific task. Each skill has a SKILL.md file describing when to invoke it, plus a references/ folder with deeper docs Claude reads as needed. Skills don't add new code execution capabilities to Claude — they steer how Claude uses tools it already has (file system, bash, web fetch, etc.).
The trove skills don't bundle their own data or query engine. They tell Claude where the published Parquet files live (on troveproject.com), how to query them with DuckDB (a CLI most engineers already have), how to read the fields, and how to cite back to the right source document. The data itself is public; the skill is the operating manual.
A skill runs inside the user's own Claude Code or Claude Desktop session. That means: their context, their conversation, their data flow — no server in the middle. It also means the user can install one skill or both, use them alongside whatever else they're working on, and audit exactly what's being read and queried.
This shape is good for reference data. The user knows their question better than a generic chatbot does, and Claude is good at translating natural-language questions into structured queries when it has the right field map.
The data is the same regardless of how you access it. The published Parquet and JSON bundles work in any environment that speaks DuckDB or HTTP: duckdb -c "SELECT … FROM read_parquet('https://troveproject.com/data/…')" from a terminal, or fetch the JSON in Python or JavaScript directly. The skill format just makes the data convenient inside a Claude session.
If you want to see the underlying data without installing a skill:
Plain-language docs for the datasets these skills query: