We don't write code like we did in 2024

At ZORC, AI-assisted development has gone from experiment to standard. Not as a gimmick — but as a tool that makes usfaster, more consistent and less prone to writing bugs.

Two tools stand out:Cursorsuch as IDE andClaudeas an AI model. Together they form the core of our development process.

Cursor: VS Code, fixed with AI built-in

What is Cursor?

Cursor is a fork of VS Code with deeply integrated AI. Your extensions, themes and keybindings work instantly. But on top of that you get:

  • Codebase-aware autocomplete— suggestions based on your entire project, not just current file
  • Chat with context— ask about your codebase and get answers with references to specific files
  • Inline editing— highlight code, describe the change, Cursor rewrites it
  • Multi-file edits— The AI ​​understands dependencies between files and changes everything needed
  • Terminal integration— run commands directly from the AI ​​chat

What does it cost?

  • Free:2,000 completions/month
  • Pro ($20/month):Unlimited completions + premium models
  • Business ($40/month):Team features + admin control

Our assessment:The Pro plan pays within the first hour of each month.

Claude: The AI ​​model we trust for code

Why Claude and not GPT?

We have tested all models extensively. Claude (Anthropic) consistently wins for our work:

Claude GPT-4o Gemini
Code quality Excellent Good Good
Context window 200K tokens 128K tokens 1M tokens
Follow instructions Best in class Good Uneven
Hallucination risk Low Means Means
Refactoring Exceptional Good OK
Explanation skills Clear, structured Verbose Short

The deciding factor:Claude follows instructions better. When we say "only change X, don't touch Y" Claude does exactly that. GPT is happy to add "improvements" we didn't ask for.

Claude Code

Anthropic CLI toolClaude Codehas become part of our daily workflow:

  • Run in the terminal, understand the whole repo
  • Can create, edit and delete files
  • Run tests and fix errors in a loop
  • Perfect for refactoring large codebases

Our AI workflow at ZORC

Phase 1: Planning (Claude in chat)

Before writing a line of code, we discuss architecture with Claude:

"We are building a booking system with Supabase. Users should be able to book appointments, pay via Stripe, and receive confirmation via email. Suggest data model, API routes and edge cases."

Claude gives us a well thought out plan with table structure, RLS policies, API endpoints and potential issues. We iterate 2-3 times before coding.

Time gain:~2 hours per project in the planning phase.

Phase 2: Scaffolding (Cursor + Claude)

With the plan ready, we use Cursor to scaffold:

  • Database migrations
  • Basic API routes
  • Component structure
  • TypeScript types

Method:We use Cursors Composer mode and input the entire plan as context. The AI ​​generates the basic structure, we review and adjust.

Phase 3: Implementation (Cursor inline + Tab)

The daily coding looks like this:

  1. Write a comment that describes what the function should do
  2. Cursor suggests implementation
  3. Tab to accept, or inline-edit to adjust
  4. Highlight complex code → "Refactor this into smaller functions"

Important principle:We never accept AI-generated code without understanding it. Each line is reviewed.

Phase 4: Review and debugging (Claude Code)

When we encounter bugs:

  1. Paste the error message into the Cursor chat
  2. Claude analyzes with full codebase context
  3. Suggest fix with explanation of root cause

For major refactoring, we run Claude Code in the terminal:

claude "Refactor the auth module to use the middleware pattern instead of individual checks in each route"

Rules we live by

1. AI writes, humans review

AI-generated code always goes through code review. Always.

2. Understand before you accept

If you don't understand what the code does, ask Claude to explain. If the explanation is not convincing, write it yourself.

3. Test everything

AI can generate tests — and we do. But we check that the tests actually test the right thing, not just that they pass.

4. Don't share sensitive data

We never send customer data, API keys or personal data to AI models. Cursor and Claude are configured with.cursorignoreand.envfiles.

5. AI does not replace architectural thinking

AI is good at implementing, bad at making strategic decisions. Data models, security architecture and business logic are still human domains.

Productivity gains in numbers

After 12 months with Cursor + Claude, we have measured:

  • 40-60% fasterfrom empty file to working feature
  • 30% fewer bugsin code review (AI catches common mistakes)
  • 2x faster onboarding— new developers understand the codebase faster with AI as a guide
  • 80% less time on boilerplate— auth, CRUD, validation generated in seconds

Where AI is still struggling

  • Complex state machines— AI loses context in deeply nested logic
  • Performance optimization— AI writes correctly but rarely optimally
  • Domain specific knowledge— industry regulations and compliance require human expertise
  • Design decisions— AI can implement design, but notdesign

Get started

  1. Install Cursor— free to try, Pro worth every penny
  2. Select Claude as the default modelin the Cursor settings
  3. Start with refactoring— mark old code, ask for improvement
  4. Write better prompts— the more context you provide, the better the results
  5. Set up.cursorignore— exclude .env, node_modules, and sensitive files

Want to see how AI-assisted development can speed up your next project?Talk to us.