Use when you have validated symmetry groups and need to design neural network architecture that respects those symmetries. Invoke when user mentions equivariant layers, G-CNN, e3nn, steerable networks, building symmetry into model, or needs architecture recommendations for specific symmetry groups. Provides architecture patterns and implementation guidance.
View on GitHubskills/equivariant-architecture-designer/SKILL.md
January 24, 2026
Select agents to install to:
npx add-skill https://github.com/lyndonkl/claude/blob/main/skills/equivariant-architecture-designer/SKILL.md -a claude-code --skill equivariant-architecture-designerInstallation paths:
.claude/skills/equivariant-architecture-designer/# Equivariant Architecture Designer ## What Is It? This skill helps you **design neural network architectures that respect identified symmetry groups**. Given a validated group specification, it recommends architecture patterns, specific libraries, and implementation strategies. **The payoff**: Equivariant architectures have fewer parameters, train faster, generalize better, and are more robust to distribution shift. ## Workflow Copy this checklist and track your progress: ``` Architecture Design Progress: - [ ] Step 1: Review group specification and requirements - [ ] Step 2: Select architecture family - [ ] Step 3: Choose specific layers and components - [ ] Step 4: Design network topology - [ ] Step 5: Select implementation library - [ ] Step 6: Create architecture specification ``` **Step 1: Review group specification and requirements** Gather the validated group specification. Confirm: which group(s) are involved, whether invariance or equivariance is needed, the data domain (images, point clouds, graphs, etc.), task type (classification, regression, generation), and any computational constraints. If group isn't specified, work with user to identify it first. **Step 2: Select architecture family** Match the symmetry group to an architecture family using [Architecture Selection Guide](#architecture-selection-guide). Key families: G-CNNs for discrete groups on grids, Steerable CNNs for continuous 2D groups, e3nn/NequIP for E(3) on point data, GNNs for permutation on graphs, DeepSets for permutation on sets. Consider trade-offs between expressiveness and efficiency. **Step 3: Choose specific layers and components** Select layer types based on [Layer Patterns](#layer-patterns). For each layer decide: convolution type (regular, group, steerable), nonlinearity (must preserve equivariance - use gated, norm-based, or tensor product), normalization (batch norm breaks equivariance - use layer norm or equivariant batch norm), pooling (for invariant outputs: us