AI-Assisted Coding Lab

In this lab you will use two AI-powered coding assistants – Claude Code and Google Antigravity – to build a complete data analytics and machine learning pipeline. The goal is to learn how these tools extend beyond simple code generation through features like hooks, skills, subagents, rules, and workflows.

Learning Objectives

• Use an AI coding assistant to plan, build, and test a full ML project
• Configure CLAUDE.md and GEMINI.md to guide AI behavior with project-specific rules
• Create hooks that automate quality checks (linting, syntax validation) on every file change
• Build custom skills that teach the AI reusable workflows (data analysis, model evaluation)
• Use subagents to decompose complex tasks into isolated, parallel work streams
• Compare how Claude Code and Google Antigravity approach the same concepts differently

Problem Statement

Your task is to build a complete ML pipeline for classifying wines into 3 classes using the UCI Wine dataset (sklearn.datasets.load_wine()). The pipeline must include:

1. Exploratory data analysis (EDA) using polars and matplotlib
2. Feature engineering and data preparation for modeling
3. XGBoost classification model training with cross-validation
4. Model evaluation and a comprehensive performance report

You will complete this task twice – once using Claude Code and once using Google Antigravity – to understand how each tool’s customization features support the development workflow. Each part includes a demo/ folder with a reference implementation on the California Housing dataset (regression), so you can see a working example without being able to copy it directly for your classification task.

Prerequisites

• Python 3.11+
• uv package manager installed
• Claude Code CLI installed (for Part 1)
• Google Antigravity IDE installed (for Part 2)

Environment Setup

# Clone the repository
git clone 
cd ai-assisted-coding

# Install uv (Python package manager)
curl -LsSf https://astral.sh/uv/install.sh | sh
source $HOME/.local/bin/env

# Install dependencies with uv
uv sync

# Verify the setup
uv run python -c "from sklearn.datasets import load_wine; print('Dataset OK')"

Lab Structure

Part 1: Claude Code

Located in part1_claude_code/. A single guided lab where you use Claude Code to build the full ML pipeline from scratch. The task naturally exercises every major feature:

• CLAUDE.md: Project-level instructions enforced on every file write
• Hooks: Automated quality checks (ruff, py_compile, test scaffolding) that fire automatically
• Skills: /analyze-data, /evaluate-model, /generate-report for reusable workflows
• Subagents: Explore, Plan, and Bash agents for task decomposition
• Slash commands: /plan for structured planning before building

See part1_claude_code/README.md for the full walkthrough.

Part 2: Google Antigravity

Located in part2_antigravity/. A single guided lab where you use Google Antigravity to build the full ML pipeline from scratch. The task naturally exercises every major feature:

• GEMINI.md + Rules: Project rules that guide agent behavior
• Workflows: /run-eda and /train-model for on-demand step-by-step instructions
• Skills: Custom data profiling capability
• Manager View: Multi-agent orchestration for parallel task execution

See part2_antigravity/README.md for the full walkthrough.

The AI-Assisted Development Workflow

Both parts follow the same development process:

1. Plan: Ask the AI assistant to create an implementation plan in a markdown file
2. Review: Read the plan, provide feedback, and request changes
3. Build: Have the AI execute the plan, writing code according to your project rules
4. Test: Hooks and automation catch issues automatically during development
5. Iterate: Review outputs, request improvements, and refine

This mirrors how professional developers work with AI coding assistants in practice.

Concept Mapping

Grading

Each part is worth 50 points (100 total).

Part 1: Claude Code (50 points)

If you have access to Claude Code, follow the full walkthrough in part1_claude_code/README.md and submit:

• The generated pipeline scripts in part1_claude_code/src/
• Output artifacts in output/
• The completed Feature Summary Checklist from the README

If you do not have access to Claude Code, complete the config-only exercise instead:

1. Read all existing configuration files (10 points): Read and understand the existing CLAUDE.md files, .claude/settings.json (hooks), .claude/skills/ (skills), and .claude/commands/plan.md (slash command). Study the demo solved code in part1_claude_code/demo/solved/ to see how the configuration influences generated code.

2. Create a CLAUDE.md for Wine classification (10 points): Write part1_claude_code/CLAUDE.md tailored to the Wine classification task. It should include dataset-specific instructions (3 wine classes, classification metrics, stratified splits) while keeping the existing coding standards.

3. Create a hook configuration (10 points): Write a .claude/settings.json that defines hooks for the Wine pipeline. Include a PostToolUse command hook for ruff and py_compile, a PostToolUse prompt hook for test file creation, and a PreToolUse command hook to block force pushes. Explain in comments or a companion markdown file what each hook does and when it fires.

4. Create skills for Wine classification (10 points): Write skill files in .claude/skills/ for at least two skills relevant to Wine classification (e.g., analyze-wine-data, evaluate-classifier). Each skill should have a SKILL.md with clear step-by-step instructions the agent would follow.

5. Create a slash command (10 points): Write a .claude/commands/plan-wine.md slash command that would generate an implementation plan for the Wine classification pipeline. The command should include a template with sections for objectives, architecture, file structure, and implementation steps.

Part 2: Google Antigravity (50 points)

If you have access to Google Antigravity, follow the full walkthrough in part2_antigravity/README.md and submit:

• The generated pipeline scripts in part2_antigravity/src/
• Output artifacts in output/
• The completed Feature Summary Checklist from the README

If you do not have access to Google Antigravity, complete the config-only exercise instead:

1. Read all existing configuration files (10 points): Read and understand .gemini/GEMINI.md, .agent/rules/code-style-guide.md, .agent/workflows/, and .agent/skills/data-profiler/. Study the demo solved code in part2_antigravity/demo/solved/.

2. Create rules for Wine classification (10 points): Write .gemini/GEMINI.md and .agent/rules/data-quality.md tailored to Wine classification. Rules should cover dataset-specific instructions, classification-appropriate metrics, and data quality checks.

3. Create workflows for Wine classification (10 points): Write .agent/workflows/run-wine-eda.md and .agent/workflows/train-wine-classifier.md with step-by-step instructions for EDA and model training on the Wine dataset. Each workflow should reference the rules and include quality check steps.

4. Create a skill for Wine classification (10 points): Write .agent/skills/wine-profiler/SKILL.md that instructs the agent to profile the Wine dataset, including class-specific statistics, feature distributions per class, and discriminative feature analysis.

5. Set up pre-commit hooks and write a comparison (10 points): Create a .pre-commit-config.yaml with ruff hooks. Write a brief comparison (300-500 words) answering: How do Claude Code hooks compare to using pre-commit + Antigravity rules? What are the advantages and limitations of each approach?

Quick Reference

# Install dependencies
uv sync

# Run a solved example (reference implementations in demo/)
uv run python part1_claude_code/demo/solved/01_eda.py
uv run python part2_antigravity/demo/solved/01_eda.py

# Lint all Python files
uv run ruff check .

# Format all Python files
uv run ruff format .

# Compile-check a file
uv run python -m py_compile part1_claude_code/demo/solved/01_eda.py