Agent Skills Hub

pytorch-lightning

PyTorch Lightning skill for scalable deep learning: automates model training, multi-GPU orchestration, data pipelines, and distributed training strategies like DDP, FSDP, and DeepSpeed.

docstring

Generate and update PyTorch-compliant function and method docstrings using reStructuredText/Sphinx conventions.

debug-distributed

Debugging guide for AReaL distributed training issues, including hangs, NCCL errors, OOM, and numerical consistency in FSDP2/TP/CP/EP.

stable-baselines3

Production-ready reinforcement learning using Stable Baselines3. Train agents, design custom environments, implement training callbacks, and optimize workflows with a scikit-learn-style API.

implementation

Structured, template-driven workflow for end-to-end feature development including coding, automated testing, verification, and session-based improvement.

langgraph

Expert LangGraph architect skill for designing stateful, multi-actor AI agent workflows with robust persistence, conditional branching, and ReAct patterns.

debug-cuda-crash

Tutorial for identifying and resolving CUDA runtime crashes using FlashInfer's API logging framework.

nvidia-resiliency-ext

Provides resiliency, health monitoring, and fault tolerance utilities for NVIDIA GPU-accelerated distributed applications, including process management and API key handling.

shipany

Official documentation skill for Shipany, an AI-powered SaaS boilerplate. Provides expert guidance on Next.js 15, Drizzle ORM, NextAuth, and payment integrations.

manimgl-best-practices

Best practices and code patterns for ManimGL (3Blue1Brown's OpenGL animation engine). Provides templates, rules for 3D/interactive scenes, camera control, and LaTeX math visualization for technical creators.

dgame-dev

Expert assistant for the DGame Unity framework, facilitating development, architecture, hotfix, and resource management within the TEngine-based ecosystem.

threejs-geometry

Three.js geometry generation: built-in shapes, BufferGeometry, vertex manipulation, custom meshes, and performance-optimized instanced rendering.