Weather Skills
⚠️ Under active development — not production ready.
These skills are an early experiment in tool composition for weather/climate
data pipelines. Interfaces, envelope schema, and skill boundaries may change
without notice. Fetchers hit real APIs and require credentials; middle-
pipeline skills have only been smoke-tested on small synthetic data. Do not
use in any automated workflow you rely on, and do not assume outputs are
scientifically validated. Expect breakage.
A set of composable Agent Skills for building
weather/climate data pipelines from an LLM-driven agent. Skills are
source-specific fetchers (ingress), generic operators that work on a shared
Zarr-based container (see ENVELOPE.md), or capabilities the
agent uses alongside pipelines.
Initiated by Rhiza Research.
Skills
Fetchers (ingress — source-specific)
| Skill | What it does |
|---|
ecmwf-fetch | ECMWF S2S ensemble precipitation forecast (cf + pf) over a --bbox (use resolve-region for a country's bbox) via ECDS → Zarr |
chirps-fetch | CHIRPS live precipitation observations → Zarr |
imerg-fetch | IMERG satellite precipitation (late release) → Zarr |
tahmo-fetch | TAHMO station observations (daily-aggregated) → Zarr |
dynamical-fetch | dynamical.org open catalog (GFS, GEFS, ECMWF IFS-ENS, AIFS, ICON-EU, MRMS, analyses) via --dataset, credential-free → Zarr |
Generic middle (operate on any envelope)
| Skill | What it does |
|---|
resolve-region | Resolve an ISO 3166-1 alpha-3 country code to a --bbox N/W/S/E (and optional boundary polygon GeoJSON) from bundled Natural Earth 1:110m boundaries |
clip-region | Subset a gridded Zarr to a --bbox N/W/S/E (use resolve-region for a country's bbox) |
aggregate-temporal | Resample along time or step into daily/weekly/dekadal/monthly windows |
deaccumulate | Convert a cumulative-since-init forecast variable (e.g. ECMWF S2S tp) into per-step diffs along the step axis |
step-to-time | Realize a forecast's step lead-time axis as wall-clock valid times (time = init + step) so it can be compared against time-based observations |
unit-convert | Convert a variable to target --to-units (e.g. precip flux kg m-2 s-1 → depth rate mm/day, via a liquid-water density bridge) |
downscale | Spatial downscaling onto a finer grid (by factor, finer resolution, or a reference grid) via --method (linear-interpolation or q-q empirical quantile mapping) |
coarsen | Coarsen or align a grid by linear interpolation onto a target (resolution, offset) — geometry only, adds no information |
rename | Rename a data variable to a new name |
concat | Join Zarr stores along a named dim (incl. new dims with coord values) |
reduce | Collapse named dims with a statistic (mean/std/min/max/sum/median) — e.g. ensemble spread as the std across number, or a time-mean baseline |
difference | Subtract one envelope from another (A − B) with inner-join alignment and broadcasting — anomalies vs a baseline, scenario-minus-historical change maps |
plot | Heatmap (optionally restricted to a --bbox and/or masked to a --mask-geojson polygon) or timeseries PNG from one dataset |
plot-compare | Side-by-side multi-panel comparison of two datasets (incl. station-vs-grid), optionally clipped to a --bbox and masked to a --mask-geojson polygon |
plot-mediogram | ECMWF-style mediogram PNG comparing a forecast ensemble against an m-climate ensemble at a single lat/lon |
Agent capabilities
Capabilities the agent uses alongside pipelines; none of them produces an
envelope output.
| Skill | What it does |
|---|
email-report | Compose an RFC 5322 .eml with attachments. Mocks SMTP — writes to disk, does not send. |
submit-feedback | Build a length-checked prefilled GitHub new-issue URL the user clicks to file feedback under their own account. Holds no token, makes no network call, creates no issue itself. |
Install
These skills live at https://github.com/rhiza-research/forecasting-skills.
There are two ways to use them.
As a CLI tool
For ad-hoc command-line use (no agent involved), install the skills as a
single forecasting-skills binary:
# One-shot, no install — list available skills
uvx --from git+https://github.com/rhiza-research/forecasting-skills forecasting-skills
# Run one
uvx --from git+https://github.com/rhiza-research/forecasting-skills forecasting-skills <skill> [args]
Or install once and invoke directly:
uv tool install git+https://github.com/rhiza-research/forecasting-skills
forecasting-skills # list
forecasting-skills <skill> [args] # run one
Each skill's PEP 723 inline dependency block is resolved by uv run --script
on each invocation, so the runner itself contributes no Python deps to the
script's runtime environment.
