AI Coding Assistants¶

Set up AI-powered coding tools to accelerate your development workflow in the lab.

Overview¶

AI coding assistants can help you write code faster, understand unfamiliar codebases, debug errors, and learn new frameworks. This guide covers the two most useful tools for lab work: GitHub Copilot (integrated into VS Code) and Claude Code (a CLI tool).

Pros and Cons¶

	Benefit
✅	Speeds up writing boilerplate and repetitive code
✅	Helps explore unfamiliar libraries and APIs
✅	Useful for debugging errors and understanding stack traces
✅	Can generate documentation and tests
❌	Generated code may contain bugs — always review
❌	May suggest outdated patterns or incorrect API usage
❌	Can be a crutch if relied on without understanding

Comparison¶

Feature	GitHub Copilot	Claude Code
Interface	VS Code extension (inline + chat)	Terminal CLI
Best For	Inline completions while coding	Multi-file edits, project-wide tasks
Context	Current file + open tabs	Entire project directory
Cost	Free with GitHub Education	API key or Claude.ai subscription
Works on OSC	Yes (VS Code Remote-SSH)	Yes (install via curl on login node)

GitHub Copilot¶

GitHub Copilot provides AI-powered code completions directly in VS Code.

Installation¶

Install the Copilot extensions in VS Code:

Open VS Code
Go to the Extensions view (Ctrl+Shift+X)
Search for and install these two extensions:
- GitHub.copilot — Inline code completions
- GitHub.copilot-chat — Chat sidebar for questions and explanations

Or install from the command line:

code --install-extension GitHub.copilot
code --install-extension GitHub.copilot-chat

Authentication¶

After installing, click the Copilot icon in the VS Code status bar
Sign in with your GitHub account
Authorize the extension when prompted

Free access with GitHub Education

Students and educators get GitHub Copilot for free through the GitHub Education Pack. Sign up with your @osu.edu email.

Usage¶

Inline Completions¶

Copilot suggests code as you type. Ghost text appears in gray:

Press Tab to accept a suggestion
Press Esc to dismiss
Press Alt+] / Alt+[ to cycle through alternatives

# Example: start typing a function and Copilot completes it
def load_dataset(path, batch_size=32):
    # Copilot suggests the body based on context

Copilot Chat¶

Open the chat sidebar with Ctrl+Shift+I (or click the Copilot icon) to ask questions:

Explain code: Select code, then ask "Explain this"
Fix errors: Paste an error message and ask for help
Generate code: Describe what you want in natural language

Useful slash commands in Copilot Chat:

Command	Description
`/explain`	Explain the selected code
`/fix`	Suggest a fix for problems in selected code
`/tests`	Generate unit tests for selected code
`/doc`	Generate documentation for selected code

Configuration¶

Add these to your VS Code settings.json (Ctrl+, → Open Settings JSON):

{
    "github.copilot.enable": {
        "*": true,
        "markdown": true,
        "yaml": true
    }
}

Copilot works over Remote-SSH

When connected to OSC via VS Code Remote-SSH, Copilot runs locally on your machine and sends completions to the remote editor. No additional setup is needed on OSC.

Claude Code (CLI)¶

Claude Code is a command-line AI assistant that can read, edit, and create files across your entire project.

What It Is¶

Claude Code is a terminal-based tool that:

Understands your full project context (reads files, searches code)
Can edit multiple files in a single session
Runs commands and interprets their output
Follows project-specific instructions via CLAUDE.md files

Installation¶

Install Claude Code using the official installer:

# Install on your local machine or OSC
curl -fsSL https://claude.ai/install.sh | bash

This installs the claude binary to ~/.local/bin/. No Node.js required.

Authentication¶

Set your Anthropic API key:

# Add to your shell profile (~/.bashrc or ~/.zshrc)
export ANTHROPIC_API_KEY="sk-ant-..."

Or authenticate interactively on first launch:

claude
# Follow the prompts to sign in

Basic Usage¶

Interactive Mode¶

Start an interactive session in your project directory:

cd ~/projects/my-research
claude

Then type natural language requests:

"Explain what train.py does"
"Add a learning rate scheduler to the training loop"
"Find all places where we load data and add error handling"

One-Shot Mode¶

Run a single prompt and print the output without entering interactive mode:

claude -p "Explain the architecture of this project"
claude -p "Generate a .gitignore for a Python ML project"

Project Configuration with CLAUDE.md¶

What It Is¶

A CLAUDE.md file gives Claude Code project-specific context — coding conventions, build commands, architecture notes, and other instructions it should follow.

Creating a CLAUDE.md¶

Place a CLAUDE.md file in your project root:

# CLAUDE.md

## Project Overview
PyTorch-based object detection pipeline for autonomous vehicle research.
Uses YOLO architecture with custom modifications for lidar-camera fusion.

## Commands
- `python train.py --config configs/default.yaml` — Train model
- `python evaluate.py --checkpoint models/best.pt` — Evaluate
- `pytest tests/` — Run tests

## Architecture
- `models/` — Model definitions (YOLO variants)
- `data/` — Dataset loaders and augmentation
- `configs/` — YAML configuration files
- `scripts/` — SLURM job scripts for OSC

## Conventions
- Use PyTorch 2.0+ features
- Type hints on all public functions
- Config files in YAML format
- SLURM scripts use `PAS1234` account

Nested CLAUDE.md Files¶

You can place additional CLAUDE.md files in subdirectories for component-specific instructions:

my-project/
├── CLAUDE.md                    # Project-wide context
├── models/
│   └── CLAUDE.md                # Model-specific conventions
├── data/
│   └── CLAUDE.md                # Data pipeline details
└── scripts/
    └── CLAUDE.md                # SLURM/OSC-specific notes

Claude Code merges instructions from all CLAUDE.md files in the path hierarchy.

Example for an ML Project¶

# CLAUDE.md

## Project Overview
Point cloud segmentation for autonomous driving using PointNet++.
Training runs on OSC Pitzer cluster with A100 GPUs.

## Commands
```bash
# Local testing
python train.py --config configs/debug.yaml --fast-dev-run

# OSC job submission
sbatch scripts/train_gpu.sh

# Evaluation
python evaluate.py --checkpoint outputs/best_model.pt --data data/test/
```

## Architecture
- `models/pointnet2.py` — PointNet++ implementation
- `data/nuscenes_loader.py` — nuScenes dataset loader
- `configs/` — Hydra YAML configs (base, debug, full)
- `scripts/` — SLURM scripts targeting Pitzer GPU partition

## Coding Conventions
- Python 3.9+, PyTorch 2.0+
- Use Hydra for configuration management
- Logging via Python `logging` module (not print)
- Save checkpoints every 10 epochs to `outputs/checkpoints/`

Settings Configuration¶

Claude Code supports three levels of settings files:

Per-Repository Settings (Not Committed)¶

.claude/settings.local.json — Personal preferences for this repo. Add to .gitignore.

{
    "permissions": {
        "allow": [
            "Bash(python *)",
            "Bash(pip install *)",
            "Bash(sbatch *)"
        ]
    }
}

Team-Shared Settings (Committed)¶

.claude/settings.json — Shared team conventions. Commit to version control.

{
    "permissions": {
        "allow": [
            "Bash(python *)",
            "Bash(pytest *)",
            "Bash(mkdocs *)"
        ]
    }
}

User-Level Global Settings¶

~/.claude/settings.json — Applies to all your projects.

{
    "permissions": {
        "deny": [
            "Bash(rm -rf *)"
        ]
    }
}

Don't commit settings.local.json

The .claude/settings.local.json file may contain personal preferences and allowed commands specific to your machine. Add .claude/settings.local.json to your .gitignore.

Best Practices¶

1. Always Review Generated Code¶

AI assistants can produce plausible-looking code that is subtly wrong. Before accepting suggestions:

✅ Read through the generated code line by line
✅ Check that API calls use correct arguments and return types
✅ Verify edge cases are handled
✅ Run tests after accepting changes

Never paste API keys, passwords, or credentials into AI chat interfaces:

❌ Don't include secrets in prompts
❌ Don't include secrets in CLAUDE.md files
✅ Use environment variables for sensitive values
✅ Add .env files to .gitignore

3. Use `.claudeignore`¶

Create a .claudeignore file (same syntax as .gitignore) to prevent Claude Code from reading sensitive or irrelevant files:

# .claudeignore
.env
*.pth
*.pt
data/raw/
wandb/
__pycache__/

4. Version Control Configuration¶

Files to commit:

✅ CLAUDE.md — Project context for the team
✅ .claude/settings.json — Shared permissions
✅ .claudeignore — Shared ignore patterns

Files to not commit:

❌ .claude/settings.local.json — Personal preferences

5. Use AI for Learning, Not Just Output¶

When an AI assistant suggests something unfamiliar:

Ask it to explain why it chose that approach
Look up the documentation for any new APIs or patterns
Understand the code before committing it to your project

Troubleshooting¶

Copilot Not Showing Suggestions¶

Check that you're signed in (look for the Copilot icon in the status bar)
Verify the extension is enabled for the current language
Check your internet connection (Copilot needs network access)
Try reloading VS Code (Ctrl+Shift+P → "Reload Window")

Claude Code "Command Not Found"¶

# Check installation
which claude

# Reinstall if needed
curl -fsSL https://claude.ai/install.sh | sh

# Ensure ~/.claude/bin is in your PATH
export PATH="$HOME/.claude/bin:$PATH"

Claude Code API Key Issues¶

# Verify your key is set
echo $ANTHROPIC_API_KEY

# If empty, add to your shell profile
echo 'export ANTHROPIC_API_KEY="sk-ant-..."' >> ~/.bashrc
source ~/.bashrc

Copilot Not Working Over Remote-SSH¶

Ensure the Copilot extension is installed locally (not on the remote)
Check that VS Code is connected to the remote (look at the bottom-left status bar)
Reload the VS Code window after reconnecting

Next Steps¶

Agent Workflows — Advanced usage: MCP servers, custom skills, hooks, knowledge management
Set up VS Code if you haven't already
Install recommended VS Code Extensions
Connect to OSC for Remote Development

AI Coding Assistants¶

Overview¶

Pros and Cons¶

Comparison¶

GitHub Copilot¶

Installation¶

Authentication¶

Usage¶

Inline Completions¶

Copilot Chat¶

Configuration¶

Claude Code (CLI)¶

What It Is¶

Installation¶

Authentication¶

Basic Usage¶

Interactive Mode¶

One-Shot Mode¶

Project Configuration with CLAUDE.md¶

What It Is¶

Creating a CLAUDE.md¶

Nested CLAUDE.md Files¶

Example for an ML Project¶

Settings Configuration¶

Per-Repository Settings (Not Committed)¶

Team-Shared Settings (Committed)¶

User-Level Global Settings¶

Best Practices¶

1. Always Review Generated Code¶

2. Don't Share Secrets¶

3. Use .claudeignore¶

4. Version Control Configuration¶

5. Use AI for Learning, Not Just Output¶

Troubleshooting¶

Copilot Not Showing Suggestions¶

Claude Code "Command Not Found"¶

Claude Code API Key Issues¶

Copilot Not Working Over Remote-SSH¶

Next Steps¶

Resources¶

3. Use `.claudeignore`¶