Social media is full of people showing off their perfect little demo apps claiming AI is revolutionary, meanwhile, AI keeps suggesting fixes for files that don’t exist or rewrites working code into broken messes.

Does that sound familiar?

Here’s the thing — the “vibe coders” are not wrong about AI being powerful. They’re just not dealing with what you’re dealing with.

You’re working on a real codebase, with real dependencies, real business logic, and real users. Real codebases are messy.

I spent ten years of my life running a development agency, and Cursor has legitimately saved me weeks of work, but only after I stopped expecting it to just “figure things out.” Now, I’m going to share with you the workflow that made Cursor work for me on complex projects.

Mental Model

Think of your AI as a brilliant intern who codes incredibly fast but has zero context about your busines or your architecture decisions.

You wouldn’t hand an intern your entire codebase and say “build the payment system.” You’d give them background, show them examples, and break down exactly what needs to happen.

Same principle applies here.

Document your patterns like someone’s life depends on it. I keep a backend-patterns.md file that explains how I structure resources — where routes go, how services work, what the data layer looks like. Every time I ask Cursor to build something backend-related, I reference this file.

Actually, maintaining these files manually got tedious, so I built somthing to handle this automatically. My tool creates files that capture your project’s architectural decisions, and keeps them updated as your codebase evolves. Personal productivity hack that’s saved me hours of documentation work.

Result = no more random architectural decisions that break everything downstream.

Plan everything before writing code. I don’t let the AI write a single line until we both understand exactly what we’re building. I usually co-write the plan with Claude or ChatGPT — what functions we need, which files get touched, potential edge cases.

This might sound tedious, but actually saves a lot of debugging time.

Show, don’t tell. Instead of explaining how something should work, point to existing code: “Build this new API endpoint, follow the same pattern as the user endpoint.” Pattern recognition is where these models actually shine.

Control scope ruthlessly. In smaller projects, you can ask it to build whole features. But as complexity grows, get more specific. One function at a time. One file at a time. The bigger the ask, the more likely it breaks something unrelated.

Maintenance

Your codebase needs to stay organized or AI starts hallucinating. Hit that reindex button in Cursor settings regularly — treat it like clearing your browser cache.

When errors happen, fix them one by one. Don’t copy-paste a wall of red terminal output. AI gets overwhelmed just like humans do.

Pro tip: Add “don’t change code randomly, ask if you’re not sure” to your prompts. This simple line has saved me countless debugging sessions.

Use Cursor Rules Files

Set up a .cursorrules file in your project root. Every time you find yourself repeating the same context in prompts — like your backend patterns or coding standards — add it to the rules file. It gets automatically included in every AI request.

You can even create file-specific rules. Your database migration files can have different guidelines than your frontend components.

Start With a Codebase Outline

Before doing any major work, have the AI create an outline of your entire project. Go file by file, documenting every class, public function, and system. Include the AI’s understanding of each piece’s purpose and any gotchas.

This used to be the most time-consuming part of my workflow, but the results were worth it. Got so frustrated with this that I ended up building my own tool to automate it. Now, it analyzes my entire project in minutes and generates comprehensive documentation, which creates a baseline that prevents the AI from building duplicate systems or breaking existing functionality.

What This Gets You

I write maybe 10% of the boilerplate I used to. Database queries with proper error handling happen in minutes instead of hours. Complex API endpoints with validation get handled while I focus on architecture decisions that actually matter.

The best part is that I can move fast on the stuff that requires actual thinking while the AI handles tedious implementation.

Your legacy codebase isn’t a disadvantage here — all that structure and business logic you’ve built up is exactly what makes AI productive. You just need to help it understand what you’ve already created.

The Real Advantage

The teams who are able to work with AI effectively on complex projects are going to have a massive advantage. That happens when they learn to give AI the context it needs to be genuinely useful.

Your competitors are still either avoiding AI completely or getting frustrated when it breaks their code.

You can be the one who figured out how to make it work, and get an advantage against your competitors who are still struggling with it.

Questions? Write to me at hi@nmn.gl (personal email) and I’ll try to help in whatever way I can.

Further Reading

P.S. Frustrated by AI forgetting existing code and making mistakes? I built something for codebases large enough that AI loses track. Loved by devs from Google, Uber, and more. Check out Giga AI.