Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/yuanruichen/MetForge itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 26, 2026MetForge is a small collection of reusable skills for atmospheric-science work. It adds domain judgment for research questions and evidence, data acquisition, scientific calculations, figure design, and atmospheric-model diagnosis while leaving general task planning, coding, tool use, and context management to the host agent. The same skills/ source works across Codex, Claude Code, and Hermes.
Quick Start
After installation, describe the scientific task naturally or name a skill explicitly. Codex uses $metforge-data, Claude Code uses /metforge:metforge-data, and Hermes uses /metforge-data; natural-language activation also works. These prompts can be copied directly:
| Scenario | Prompt |
|---|---|
| Develop a research project | Use $metforge-project to develop spcz-moist-margin through focused questions, then maintain doc/core/research/spcz-moist-margin/proj.md and log.md. Start from my notes and ask one key question at a time. |
| Download reanalysis data | Use $metforge-data to download ERA5 hourly u and v at 850 hPa for 10°S–10°N, 80°E–180°E during DJFM 2000–2020. Validate one month before scaling up. |
| Inspect and prepare local data | Use $metforge-data to inspect these NetCDF files, verify coordinates, units, calendar, and missing values, then design a memory-safe xarray preprocessing script. |
| Calculate an atmospheric index | Use $metforge-analysis to calculate this SPCZ intensity and orientation index. State the exact domain, weighting, baseline, units, and validation before running it. |
| Trend, filtering, and statistics | Use $metforge-analysis to detrend and 20–100-day filter this continuous daily field, then calculate EOFs, regression maps, and lag composites with serial-dependence-aware uncertainty. |
| Design an atmospheric figure | Use $metforge-figure to turn these MJO composite fields into a clear longitude–time figure with consistent scales, units, and propagation guides. |
| Audit an existing figure | Use $metforge-figure to audit this multi-panel map for scientific comparability, color normalization, labels, projection artifacts, and export quality. |
| Diagnose a dynamical-core test | Use $metforge-model-diagnose to determine why RMS error grows in this balanced-flow test and propose the smallest experiment that separates spatial imbalance from a growing computational mode. |
| Combine analysis and visualization | Use $metforge-analysis to calculate the diagnostic fields and statistics, then use $metforge-figure only for the panels needed to support the conclusion. |
Skills
| Skill | Use it for | Example |
|---|---|---|
metforge-project | Clarify research questions and hypotheses through dialogue; create, resume, and update proj.md and log.md | “Use $metforge-project to turn this research idea into a concise project core.” |
metforge-data | Find, download, subset, validate, and document atmospheric datasets | “Use $metforge-data to download ERA5 hourly winds for this domain.” |
metforge-analysis | Calculate indices, meteorological diagnostics, trends, filters, EOFs, regression, and composites | “Use $metforge-analysis to calculate this circulation index and its trend.” |
metforge-figure | Design, create, revise, and audit atmospheric-science figures | “Use $metforge-figure to turn these balance-test outputs into a publication-ready comparison.” |
metforge-model-diagnose | Diagnose dynamical-core tests, numerical experiments, balance, conservation, convergence, and error growth | “Use $metforge-model-diagnose to determine why the balanced-flow experiment drifts.” |
Each directory under skills/ is an independent installable unit. The skills may cooperate, but none requires a central orchestrator.
Research project documents
metforge-project develops the scientific design through a short, adaptive interview. It separates literature, observations, hypotheses, and interpretation. Existing project notes are reused; unresolved choices remain visible until the researcher settles them.
Use $metforge-project to start a project named spcz-moist-margin from these notes. Ask one key question at a time.
Use $metforge-project to resume spcz-moist-margin from its documents and continue the next open question.
Use $metforge-project to update spcz-moist-margin from this new evidence: append log.md, reassess H1, and revise the affected parts of proj.md.
The documents belong in the active research repository at doc/core/research/<projname>/: proj.md holds the concise current core, and log.md retains dated evidence and decisions. Meaningful checkpoints save the draft and next question for later sessions. Confirmation of the plan does not verify its hypotheses; initialization records no experimental results unless actual evidence was provided. Detailed execution records stay in the project's existing provenance location.
Install
Codex
Clone the repository and sync all skills:
git clone https://github.com/yuanruichen/MetForge.git
cd MetForge
scripts/install-codex-skills.sh
Verify an existing installation without changing it:
scripts/install-codex-skills.sh --check
By default the installer uses ${CODEX_HOME:-$HOME/.codex}/skills. Use --dest PATH to install elsewhere.
Update an existing Codex installation
Run these commands from your existing MetForge checkout on main:
git pull --ff-only origin main
scripts/install-codex-skills.sh
scripts/install-codex-skills.sh --check
This fetches the latest source, syncs every skill (including newly added skills), and verifies the installed copies. If you originally used --dest PATH, pass the same option to both installer commands. Keep custom changes in the source checkout: syncing replaces managed installed copies. If Git cannot fast-forward because of local changes, reconcile them before retrying; the update does not reset your work. If a new skill is not visible, open a new agent conversation.
Claude Code
Register this repository as a marketplace, then install the MetForge plugin:
/plugin marketplace add yuanruichen/MetForge
/plugin install metforge@metforge
Claude Code discovers all five skills from the shared skills/ directory. They are available as /metforge:metforge-project, /metforge:metforge-data, /metforge:metforge-analysis, /metforge:metforge-figure, and /metforge:metforge-model-diagnose.
Hermes
Add MetForge as a public Skills Hub tap:
hermes skills tap add yuanruichen/MetForge
Install the skills directly from GitHub:
hermes skills install yuanruichen/MetForge/skills/metforge-project
hermes skills install yuanruichen/MetForge/skills/metforge-data
hermes skills install yuanruichen/MetForge/skills/metforge-analysis
hermes skills install yuanruichen/MetForge/skills/metforge-figure
hermes skills install yuanruichen/MetForge/skills/metforge-model-diagnose
Hermes treats third-party taps as community sources and security-scans skills during installation.
Design principles
- Domain judgment over agent orchestration.
- A small, testable result before a costly run.
- Explicit scientific contracts: variables, units, coordinates, hypotheses, metrics, and evidence.
- Choose direct execution or Slurm from the actual runtime, data volume, and cost instead of assuming one environment.
- Append a processing record after every completed download, calculation, model run, or render stage.
- Raw inputs remain immutable; derived products remain reproducible.
- Figures are scientific arguments and are inspected after export.
- Claims are separated into evidence, interpretation, and unresolved uncertainty.
Codex, Claude Code, and Hermes share the same skill directories; platform-specific files contain distribution metadata only.