Using Cursor AI for iOS app development

Lately if feels like there has been an explosion of AI tools, each with its own specific use-case, or twist, and honestly it’s freaking hard to keep up with what’s actually useful, specially for more “niche” apps like mobile native, as seems that every new thing that comes out is made for the web dev world.

For the past year I’ve been sticking mostly to ChatGPT, and the recent desktop app update that integrates directly with Xcode has been a game-changer.

This got me thinking: what other tools am I missing out on? I’m definitely not using AI to its potential (not even 1%), so this week I’ve been exploring several tools, in particular Claude, Midjourney, and Cursor. I haven’t had much time to dive deep, but I can already see their potential for my projects (and work!).

What is Cursor?

Cursor is basically an IDE, a code editor has AI models built-in. I use it with Claude 3.5 sonnet just because it was the default one, but there are more to choose from. You can go ahead to cursor.com and download their free trial.

Working with Cursor and Xcode

I’ll be honest: It hasn’t been the most straight-forward experience, I’ve tried a few times before and the results were always catastrophic. Issues like non-compiling code, fabricated methods, odd design patterns, unit tests not testing what should be testing … you call it.

After some tinkering I found a setup that works well for small proof-of-concept projects, and discovered that most of my initial issues were my fault due not setting up Cursor correctly for iOS development.

Initial setup: 

• Start your project in Xcode, then, and only then, open the same folder in Cursor. This ensures Cursor has access to the entire workspace.

• Provide relevant docs: I think one of the key factors to make it work is to make Cursor aware of which documentation should be following, in this case, I’ve added docs to Swift, SwiftUI, Apple’s design tips, and the Human Interface Guidelines. Now when you prompt to code, it will take those into account:

  

  Docs are key to make it work properly

Workflow:

• Save changes manually: As with VScode, if you apply any change in Cursor you still need to cmd+s to save changes. Only then these will appear in Xcode. It’s easy to forget when Xcode users are not used to this, as well as for multi-file changes, so be sure to take a look to the IDE and save what’s unsaved!

• The cursor composer can be opened with cmd+shit+i. You will use this for multi-file codebase changes. While you can simply use the AI chat for same-file changes, if you need Cursor to take into account the project as a whole you can tag it with @codebase

Sync:

Do you still need to use Xcode for iOS apps? Yep, you still need Xcode to manage the project, compile it, run the simulator, etc, …

Prompting:

When giving instructions to Cursor, clarity is key. A scrappy paragraph won’t cut it. Be explicit with project specifications, as AI won’t assume anything or will give the “wrong” implementation. In my case I started writing, leave it cooking for a few minutes, and come back and add clarifications. Once done, I just made an specs.md file, and added it to the root folder, this is useful so you can always reference it when chatting with the AI.

In this example I wanted to try and generate a clicker game using SpriteKit, and Cursor was able to create the proof of concept for the app in only one prompt:

As you can see the minimum functionality is there, and compiles without any issue.

This approach worked great to translate project specs directly into code without writing a single line myself. There will some some design changes to make so we can effectively grow the code base, but has saved me a day of making the initial prototype.

Do you still need to know how to code? You better, specially for anything more complicated, or if requires a minimum of iteration.

It’s also easy to lose track of what AI is doing when using Composer and multi-file editing, so it’s dangerous to just agree to whatever code spits out.

You will need to review it and make a judgement call.

As a developer, both professionally and as a hobbyist, I’m genuinely excited about the possibilities AI tools are bringing. At the same time, I can’t shake the feeling that my job as it is now will become obsolete sooner than later.

It’s not about the job disappearing entirely, but about it transforming into something fundamentally different. In the future, we might spend far less time writing code and more time defining specs, guiding decisions, and shaping what the AI builds for us.

The challenge comes in knowing how to adapt to this shift.

PD: I’m thinking I’ll continue this project trying to use AI only (when possible) as a learning exercise. Now that Cursor has created the main scaffolding for the app, I’ll use Midjourney next for the graphics/UI. Let’s see where it goes!