Setup & Installation
Or with the ClawHub CLI, for registry-managed skill folders outside a full OpenClaw workspace:
Version History
Major update: The skill now generates full runnable code and training pipelines for S2S AI weather models. - Renamed from "s2s-forecasting-expert" to "s2s-model-builder" to reflect code generation capabilities. - Provides end-to-end PyTorch code for model architecture, training, inference, and evaluation. - Supports FuXi-style, FengWu-style, and AIFS-inspired S2S forecasting models. - Implements CRPS loss, probabilistic outputs, distributed training, and ERA5-style data workflows. - Maintains strict security: all code runs locally, with no external API calls or data transmissions.
What This Skill Does
Generates runnable PyTorch code for Subseasonal-to-Seasonal (S2S) weather forecasting systems. Covers model architectures inspired by FuXi, FengWu, and AIFS, plus training loops, CRPS loss, and evaluation scripts. All code runs locally with no external API calls.
Provides domain-specific architecture patterns and loss functions for S2S forecasting that general code assistants rarely get right without extensive prompting.
When to use it
- Building a FuXi-style transformer for 30-day Z500 forecasting
- Implementing CRPS loss for probabilistic ensemble outputs
- Scaffolding an ERA5 data preprocessing and training pipeline
- Setting up distributed multi-GPU training for global grid models
- Generating evaluation scripts comparing model skill against climatology
Example Workflow
Here's how your AI assistant might use this skill in practice.
User asks: Generate a FuXi-style transformer in PyTorch for 30-day Z500 forecasting with CRPS loss
- 1Designs a 3D spatiotemporal transformer architecture with global grid attention
- 2Generates multi-variable input pipeline handling Z500, T2M, and wind fields
- 3Implements lead-time conditioned decoder with ensemble output heads
- 4Creates CRPS loss function in both Gaussian and ensemble forms
- 5Outputs training loop with mixed precision and checkpoint saving
Complete PyTorch module files for model, loss, and training loop ready to run locally