parquetry skills
Two skills for AI agents working with GeoParquet:
reference-data answers spatial questions with real reference data.write-geoparquet writes optimized GeoParquet the way these datasets
are written (see below).
Two answers the skill gave, mapped from the same files: 95,264 buildings in
Eastern Massachusetts touch a 100-year flood zone (FEMA NFHL and
OpenStreetMap, Boston waterfront shown), and Hérault is 49% forest and
semi-natural land while vineyards are its largest single class (CORINE Land
Cover 2018 clipped to FAO GAUL).
Answering spatial questions
The reference-data skill lets an agent answer questions with real data,
not general knowledge:
- What is the predominant land use in Hérault, France?
- How many buildings could be hit by a 100-year flood in Eastern Massachusetts?
- How many km of 380 kV line cross Lombardy?
- Which substations are within 5 km of this site?
- Add the country, district and flood zone to my list of 300 sites.
- Which dams are on the Colorado River and its tributaries?
- Which congressional and state legislative districts do I vote in?
- How many people live within 5 km of my sites, and what do they earn?
- Which natural hazards threaten this site most?
- Is this site on protected land, and who manages it?
- Geocode these addresses and tell me which are in the flood zone.
The agent turns the question into a spatial query over eighteen open datasets
published by Geomermaids on parquetry,
runs it with DuckDB straight from the URLs, and answers with the number,
the definitions it chose, the coverage, the data dates and the credits.
There is no server, no API key and no download step: DuckDB reads only the
parts of the GeoParquet files the question needs.
| Dataset | Covers | Answers |
|---|
| FAO GAUL 2024 | World | Countries, regions, districts: the areas questions are about |
| GMWID | World | Power, telecoms, oil and gas, water infrastructure |
| FEMA NFHL | United States | 100-year and 500-year flood zones, updated daily |
| CORINE Land Cover 2018 | Europe | Land cover and land use, 44 classes, and on H3 cells for large areas |
| Geoconnex | United States | Rivers and their network, dams, gages, watersheds, aquifers, water monitoring sites |
| Census TIGER/Line 2026 | United States | Congressional and state legislative districts of the 2026 elections |
| Census ACS 2020-2024 | United States | Population, income, poverty, housing, education and commuting by state, county, place, tract, block group and ZIP code area |
| FEMA National Risk Index | United States | Risk of 18 natural hazards by tract and county: expected loss, social vulnerability, resilience |
| PAD-US 4.1 | United States | Protected and conservation lands: owner, manager, designation, GAP status, public access |
| National Address Database | United States | 99 million address points for geocoding, with a coverage map |
| Gazetteer | US, World | Place names to coordinates: Census Gazetteer 2026 and GeoNames |
| WorldPop 2025 |
Writing GeoParquet
The write-geoparquet skill turns any vector source DuckDB reads into a
GeoParquet file that is fast over HTTP: GeoParquet 2.0 with native
GEOMETRY, a bbox covering column, Hilbert order, row groups capped by bytes
(DuckDB will not write them under 2,048 rows, so pyarrow and geoarrow cut
them), optional subdivision of heavy polygons, partitioning with a
manifest, and a verification of every file. Each rule comes with the
measurement behind it, from building these datasets. It also covers the
PMTiles archive for maps (tiler settings, styling in the tiles, H3 cells
for layers too dense to tile) and catalog extents that cross the
antimeridian.
uv run skills/write-geoparquet/scripts/write_geoparquet.py zones.gpkg zones.parquet --make-valid --subdivide 100
Install
Claude Code
/plugin marketplace add gsueur/parquetry-skills
/plugin install parquetry@geomermaids
The skills load when a task needs them. You can also call them by name:
/parquetry:reference-data, /parquetry:write-geoparquet.
Other agents
The skills are plain Markdown in skills/: each SKILL.md is the method,
reference/ holds the details. Point any agent that can run code at those
files.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
