Writing Code with Cursor: Fast, Fun… and Occasionally Fragile

AI-assisted coding feels a bit like having Jarvis from Iron Man in your IDE — until it suddenly forgets what suit you’re building. After spending time building with Cursor, here’s my honest take.

🚀 Where It Shines

Cursor is incredibly good at straightforward tasks:

  • UI components
  • CRUD operations
  • REST APIs
  • Boilerplate setup

In these areas, it’s like autopilot. You describe what you want, and it delivers something usable in seconds. Productivity goes up. Features roll out faster. The dopamine hits are real.

For “vibe coding” complete features, it often works surprisingly well. You prompt, it builds, and suddenly you have something functional. It feels like coding on fast-forward.

🧠 Where It Struggles

The cracks start to show with complex business logic and large codebases.

  • It fixes the issue you prompt.
  • It handles the edge case you mention.
  • But it may quietly introduce other unintended issues.

Not literally a thousand bugs… but enough to make you nervous.

The real problem is context. As the feature grows, the AI’s understanding can drift. And when debugging starts, things get tricky. Fixing bugs becomes more vibe coding. That vibe coding introduces new bugs. And suddenly you’re in a loop that feels suspiciously like a plot twist from Black Mirror.

🛠 My Ideal Workflow

Here’s what’s working for me:

1. Agent Mode → For simple tasks

  • UI
  • CRUD
  • APIs
  • Mechanical refactors

Let it move fast.

2. Ask Mode → For business logic

  • Discuss the approach first.
  • Keep prompting until the structure makes sense.
  • Read the generated code carefully.
  • Make adjustments yourself.

This forces you to stay in context. You understand the system instead of outsourcing the thinking.

When everything looks good, either:

  • Copy-paste the parts you trust, or
  • Let Cursor implement it — but with supervision.

I haven’t explored Plan Mode yet. That’ll be a separate experiment.

🧩 Final Thoughts

I still prefer to read and write code myself.

If I know exactly what I want and how it should look, AI is an excellent accelerator. If I give vague prompts and let it “figure it out,” I’m usually the one who ends up figuring things out later — during debugging.

AI is a powerful co-pilot. But I’m not ready to hand over the steering wheel.

P.S: I had AI write this blog. I thought out loud and AI structured the content to a blog. Is it any good? Let me know!