Burn Expert Plugin

Comprehensive Claude Code plugin for the Burn deep learning framework.

Features

• Domain Skills: Tensors, modules, training, ONNX import, backends, contributing
• Workflow Commands: /burn-new, /burn-train, /burn-import, /burn-smoke
• Specialized Agents: Debugger, code reviewer, ONNX surgeon
• Self-Verification: Auto-format and compile check via hooks
• Evidence-Based: Searches documentation before asserting API behavior

Setup

1. Install llmx-mcp

The plugin uses llmx for documentation search. Build and install the MCP server:

# Clone and build llmx
git clone https://github.com/zackees/llmx.git
cd llmx/ingestor-core
cargo build --release --bin llmx-mcp --features mcp

# Add to PATH (or symlink)
cp target/release/llmx-mcp ~/.local/bin/

2. Pre-index Burn Documentation (Optional)

The plugin includes pre-built indexes. To rebuild:

export LLMX_STORAGE_DIR=/path/to/burn-plugin/data/indexes
llmx-mcp  # Then send index commands via JSON-RPC

3. Install Plugin

# Option A: Test with plugin directory flag
claude --plugin-dir /path/to/burn-plugin

# Option B: Create a marketplace wrapper and install
# See Claude Code plugin documentation for marketplace setup

Structure

burn-expert/
├── .claude-plugin/plugin.json    # Plugin manifest
├── .mcp.json                     # MCP server config (llmx)
├── commands/                     # Slash commands
├── skills/                       # Domain knowledge & specialized skills
│   ├── burn-router/              # Central routing
│   ├── burn-app-dev/             # Tensors, modules, config
│   ├── burn-training/            # Learner, metrics, loops
│   ├── burn-onnx/                # Model import
│   ├── burn-backends/            # Backend selection
│   ├── burn-contrib/             # Contributing to Burn
│   ├── burn-debugger/            # Error diagnosis
│   ├── burn-reviewer/            # Code review
│   └── burn-onnx-surgeon/        # ONNX import fixes
├── hooks/                        # Auto-verification
└── data/                         # Documentation corpus
    ├── burn-book/
    ├── burn-contributor-book/
    └── indexes/                  # llmx search indexes

Usage

Commands

/burn-new my-project wgpu    # Scaffold new Burn project
/burn-train mnist            # Add MNIST training loop
/burn-import model.onnx      # Import ONNX model
/burn-smoke                  # Format, check, test

Skills (auto-activate)

Ask about:

• Tensor operations, shapes, types
• Module definition and configuration
• Training loops and metrics
• ONNX/PyTorch model import
• Backend selection and custom kernels
• Contributing to Burn

Specialized Skills

• burn-debugger: Diagnose runtime errors, shape mismatches
• burn-reviewer: Code review for Burn idioms and best practices
• burn-onnx-surgeon: Fix ONNX import failures

Design Principles

1. Version-aware: Detects project Burn version, warns on drift
2. Evidence-seeking: Cites documentation chunks, not assumptions
3. Self-verifying: cargo check after every code generation

Corpus Version

Documentation from Burn 0.16.x (burn-book + burn-contributor-book).