S2S Forecasting Expert (FuXi, FengWu, AIFS)

End-to-end builder for AI-based Subseasonal-to-Seasonal (S2S) forecasting systems. Generates runnable PyTorch code for FuXi-style, FengWu-style, and AIFS-ins...

Security Scan

This instruction-only skill is internally consistent with its stated purpose: it generates local PyTorch scaffolds for S2S forecasting and does not request credentials, install software, or call external endpoints.

Purpose & Capability

The name and description (S2S model builder for FuXi/FengWu/AIFS-style systems) match the SKILL.md content: generation of model architectures, training loops, CRPS implementations, evaluation, and inference code. No unrelated capabilities or credentials are requested.

Instruction Scope

SKILL.md is an instruction-only manifest that directs the agent to generate code and scaffolds. It does not instruct the agent to read arbitrary host files, access environment variables, or send data externally; it explicitly states no external API calls or automatic dataset downloads.

Install Mechanism

There is no install spec and no code files executed by the platform. Being instruction-only means nothing is downloaded or written by the skill itself—lowest-risk install posture.

Credentials

The skill declares no required environment variables, credentials, or config paths. This is proportional to an LLM-based code-generator that produces local training/inference code.

Persistence & Privilege

always is false and the skill does not request persistent or system-wide configuration changes. Autonomous invocation is allowed by default but is not combined with other high-risk behaviors.

Assessment

This skill appears coherent and low-risk, but exercise normal caution: review any generated code before running it (look for dataset download commands, remote URLs, or hard-coded credentials), run unknown training scripts in a controlled environment (container/VM) if possible, verify you comply with dataset licenses such as ERA5, and disable the skill if you do not want it to be auto-invoked for related queries. If you plan to run distributed or multi-GPU scripts, inspect scheduler/launcher commands to avoid unintentionally interacting with cluster resources or credentials.

S2S Model Builder (Subseasonal-to-Seasonal Forecasting)

This skill actively helps you design, implement, and train S2S forecasting models from scratch.

It generates:

• PyTorch model architectures
• Training loops
• CRPS loss implementations
• Data preprocessing pipelines (ERA5-style)
• Evaluation scripts
• Multi-GPU training configurations
• Inference pipelines

Supported paradigms include:

• FuXi-style transformer architectures
• FengWu-style Earth system transformers
• AIFS-inspired probabilistic models
• Ensemble neural forecasting
• Multi-lead-time forecasting heads

---

What This Skill Can Build

1. Model Architecture Code

• 3D spatiotemporal transformers
• Global grid attention models
• Multi-variable input pipelines (Z500, T2M, winds, SST)
• Lead-time conditioned decoders
• Ensemble output heads

2. Training Infrastructure

• PyTorch training loops
• Distributed training (FSDP-ready structure)
• Mixed precision support
• Gradient accumulation
• Checkpoint saving

3. Probabilistic Forecasting

• CRPS loss (Gaussian & ensemble forms)
• Quantile regression heads
• Spread-skill diagnostics
• Reliability calibration utilities

4. Evaluation Code

• CRPS computation
• ACC metric implementation
• RMSE across forecast horizons
• Skill vs climatology baseline

5. Deployment-Ready Inference

• Batched inference scripts
• Memory-optimized forward passes
• Model export patterns

---

Example Prompts

• “Generate a FuXi-style transformer in PyTorch for 30-day Z500 forecasting.”
• “Build a CRPS loss function for ensemble S2S outputs.”
• “Create a full ERA5 training pipeline scaffold.”
• “Design a multi-lead-time S2S forecasting head.”
• “Implement distributed training for global 1° resolution data.”

---

External Endpoints

This skill does not call external APIs.

Endpoint	Purpose	Data Sent
None	N/A	None

All generated code runs locally within the user’s environment.

---

Security & Privacy

• No external API calls
• No automatic dataset downloads
• No remote execution
• No hidden scripts
• All code is generated transparently

Users are responsible for lawful dataset usage (e.g., ERA5 licensing).

---

Model Invocation Note

This skill may be automatically invoked when user queries involve:

• Building S2S models
• FuXi / FengWu / AIFS implementations
• CRPS training
• AI weather model architecture
• ERA5 training pipelines

Users may opt out by disabling the skill.

---

Trust Statement

By using this skill, you acknowledge it generates code for AI-based climate forecasting systems. No data is transmitted externally. All execution occurs within your own environment.

---

Version

v1.0.0
Last updated: Feb 16, 2026