Weather Skills

⚠️ Under active development — not production ready.

These skills are an early experiment in tool composition for weather/climate data pipelines. Interfaces, dataset 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.

Small tools an AI agent can run to fetch weather data and transform it. Each skill is a command-line script. They share one Zarr file format so outputs of one skill can feed into the next — see STANDARD_DATASET.md.

Built as Agent Skills by Rhiza Research.

Skills

Fetchers (ingress — source-specific)

Prefer dynamical-fetch when the dynamical.org catalog has the dataset (GFS, GEFS, ECMWF IFS-ENS, AIFS, Google WeatherNext 2, ICON-EU, analyses, IMERG, CHIRPS). Use a credentialed or source-specific fetcher only when it does not.

Generic middle (operate on any standard dataset)

Agent capabilities

Capabilities the agent uses alongside pipelines; none of them produces a dataset output.

Install

These skills live at . There are three ways to use them.

As a Claude Code plugin

Install the plugin once — add the marketplace, then install the plugin:

claude plugin marketplace add rhiza-research/weather-skills
claude plugin install rhiza-forecasting@weather-skills

Then run the bundled forecaster agent. The --allowedTools flag pre-approves the plugin’s skills for the session so a multi-step pipeline runs end to end without a prompt at each step:

claude --agent rhiza-forecasting:forecaster --allowedTools "Skill(rhiza-forecasting:*)"

Fetchers still need their credentials in the environment; see each skill’s compatibility: frontmatter for what it reads (for example an EXAMPLE_API_KEY).

To make the approval permanent instead of passing the flag every time, add the same rule to permissions.allow in a settings file (.claude/settings.json for one project, ~/.claude/settings.json for every project):

{
  "permissions": {
    "allow": [
      "Skill(rhiza-forecasting:*)"
    ]
  }
}

With that rule in place, plain claude --agent rhiza-forecasting:forecaster works without the flag. This lets the plugin’s skills run unprompted — including reaching the network and writing output files — so add an ask or deny rule instead if you want to be prompted for some or all of them.

As a CLI tool

For ad-hoc command-line use (no agent involved), install the skills as a single weather-skills binary:

# One-shot, no install — list available skills
uvx --from git+https://github.com/rhiza-research/weather-skills weather-skills

# Run one
uvx --from git+https://github.com/rhiza-research/weather-skills weather-skills  [args]

Or install once and invoke directly:

uv tool install git+https://github.com/rhiza-research/weather-skills
weather-skills                          # list
weather-skills  [args]           # run one

Each skill’s PEP 723 inline dependency block is resolved by uv run on each invocation, so the runner itself contributes no Python deps to the script’s runtime environment.

As agent skills

For use by an LLM agent (Claude Code, etc.), install the SKILL.md files into your project with skillkit — no local install needed, npx runs the latest skillkit on demand (add @latest to always pull the newest):

# List what skillkit discovers in the repo
npx skillkit install rhiza-research/weather-skills --list

# Install all skills to the current project
npx skillkit install rhiza-research/weather-skills --all --yes

# Install globally so any project can use them
npx skillkit install rhiza-research/weather-skills --all --yes --global

# Target a specific agent (otherwise skillkit installs for every agent it detects)
npx skillkit install rhiza-research/weather-skills --all --yes --agent claude-code

# Install just a subset
npx skillkit install rhiza-research/weather-skills --skill=ecmwf-fetch
npx skillkit install rhiza-research/weather-skills --skills=clip-region,difference

# Overwrite an existing install
npx skillkit install rhiza-research/weather-skills --all --yes --force

Pin in a manifest for team / reproducible use:

npx skillkit manifest init
npx skillkit manifest add rhiza-research/weather-skills
npx skillkit manifest install

See npx skillkit install --help for more flags.

Composition pattern

Middle-pipeline skills are designed to chain. Example — a weekly forecast and the matching observations for one country, using the weather-skills CLI from the Install section above:

# Kenya bbox (N/W/S/E) from resolve-region
weather-skills resolve-region KEN

# ECMWF 46-day ensemble forecast, weekly totals
weather-skills dynamical-fetch \
    --dataset ecmwf-ifs-ens-forecast-46-day-daily-1-5-degree \
    --date 2026-09-14 --bbox 5.5/33.9/-4.7/41.9 -v tp \
    --output /tmp/ecmwf.zarr
weather-skills aggregate-temporal \
    --input /tmp/ecmwf.zarr --period weekly --output /tmp/ecmwf_weekly.zarr
weather-skills convert-to-totals \
    --input /tmp/ecmwf_weekly.zarr --output /tmp/ecmwf_weekly_totals.zarr

# CHIRPS observations for the same weeks
weather-skills chirps-fetch \
    --start-time 2026-09-14 --end-time 2026-09-20 --bbox 5.5/33.9/-4.7/41.9 \
    --output /tmp/chirps.zarr
weather-skills aggregate-temporal \
    --input /tmp/chirps.zarr --period weekly --output /tmp/chirps_weekly.zarr
weather-skills convert-to-totals \
    --input /tmp/chirps_weekly.zarr --output /tmp/chirps_weekly_totals.zarr

In practice a user just states the goal in natural language and the agent picks and composes skills from this set.

Standard dataset contract

Skills share a simple Zarr contract: fixed dimension names (latitude, longitude, time, step, number, station_id, …) and short types (spatial, observations, forecast, vertical_forecast, point_obs, …). See STANDARD_DATASET.md. Fetchers write that shape; other skills only depend on dims, coords, data variables, and weather_skills_* attrs — not on per-variable encoding.

CLI flag conventions

Each skill declares its CLI through the @weather_skill decorator from weather_skills_core, so common parameters (--input / -o, --bbox, --start-time / --end-time, etc.) mean the same thing wherever they appear. See CONVENTIONS.md for the full mapping of concept → canonical flag.

Credentials

Fetchers read credentials from environment variables (or .netrc where supported by the underlying client). Nothing is hardcoded. See each skill’s compatibility: frontmatter for the specific vars required.

Contributing

See CONTRIBUTING.md for the publishing model (main is the consumer-facing branch — every merge is a release), the PR workflow, and the version-bump conventions (per-skill metadata.version in SKILL.md, driven by release: major / release: minor PR labels, with patch as the default).

License

MIT. See LICENSE.