Code-Off: Claude Code vs Cursor

AI Coding Assistant Comparison - Expense Tracking Application

Presentation Materials

This repository contains technical analysis and documentation comparing two implementations of the same expense tracking application, built with different AI coding assistants.

Documents

• COMPARISON.md - Comprehensive side-by-side comparison

  • Architecture differences
  • Code quality metrics
  • Development time analysis
  • Prompt engineering lessons
  • Production readiness assessment
  • Cost-benefit analysis

• ARCHITECTURE_CLAUDE.md - budget_claude (Flask + Claude Code)

  • Monolithic architecture
  • System diagrams
  • Performance metrics
  • Design decisions

• ARCHITECTURE_CURSOR.md - budget_cursor (FastAPI + Cursor)

  • Modular architecture
  • Testing infrastructure
  • Validation patterns
  • Future improvements

The Experiment

Objective: Build an AI-powered expense tracking app with transaction categorization using local LLMs (Ollama + llama3.1:8b)

Implementations:

• budget_claude: Built with Claude Code (Anthropic Claude Sonnet 4.5)
• budget_cursor: Built with Cursor (using Claude models)

Key Findings:

• Claude Code: 40% faster to MVP, favored rapid iteration (monolithic structure)
• Cursor: Generated more production-ready code (modular, tested, typed)
• Both successfully integrated local LLM with comparable performance
• Local Ollama: $0/month vs cloud APIs $20-50/month
• Accuracy: 85-90% with historical context learning

Presentation Focus

For builder community interested in:

• AI-assisted development trade-offs
• Local LLM deployment strategies
• Code quality vs development speed
• Practical prompt engineering lessons
• Production readiness assessment

Sample Data

The data/ directory contains anonymized sample CSV files in multiple formats for testing both applications:

• simple_transactions.csv - Basic 3-column format (14 rows)
• credit_card_transactions.csv - Credit card statement format (15 rows)
• chase_transactions.csv - Chase bank export format (13 rows)
• bank_checking_transactions.csv - Bank checking format (15 rows)

See data/README.md for detailed format descriptions.

Source Repositories

• budget_claude - Flask + Claude Code implementation
• budget_cursor - FastAPI + Cursor implementation
• budget_tracing - Shared Langfuse monitoring server

Technical Stack

Both Apps:

• Frontend: React
• Backend: Python (Flask vs FastAPI)
• LLM: Ollama with llama3.1:8b (local)
• Monitoring: Langfuse (optional)
• Container: Docker + Docker Compose

Key Difference: Architectural philosophy influenced by AI assistant choice