How I built a SaaS app in 30 minutes

I had a post of me building software with AI blow up on LinkedIn.

Why?

Because people were shocked at what’s possible when you know how to use AI effectively

Today, I’ll show you how I built jsonl-builder.com in 30 minutes. I’ll go over tips on prompting, setting up a project in Cursor, and choosing the right AI models

We’re entering an era where software is expected to be built in a fraction of the time it once took.

So to make sure you can replicate what I did, today we’ll go through:

• What software and which AI models I use for coding

• How I set up my development environment for success

• The prompt that built the whole app

What software and AI models to use

I have tried a ton of “code editors” with AI functionality. The following are some of them:

• VS Code with Github Copilot

• Cursor

• Windsurf

• Zed

• v0

• Lovable

I’ve found myself sticking with Cursor.so. It’s the clear frontrunner for AI-assisted coding, and with their recent $105M Series B, I don’t see that changing anytime soon.

My LLM of choice is mostly Claude 3.7 Sonnet. However, it’s not perfect for all cases, and I sometimes switch between models. I’ll dive deeper into this in an upcoming newsletter issue 🚀

Setting up your environment for success

Creating a clear project description with information about what you want to build is crucial for the AI to understand your vision.

Most code editors allow for some sort of ruleset you can write that tells the AI how to act.

Cursor has .cursor/rules
VS code has .github/copilot-instructions.md

These files act as a description of how the AI should write code, what tech it should use, and where it can find the relevant pieces of your project.

You can also describe what your project does.

This way, the AI will keep your goals in mind as it adds new functionality.

Here’s the exact cursor rules I used to launch jsonl-builder.com in 30 minutes:

First I describe what the AI is supposed to do.

You are an AI building a web app that helps users generate .jsonl files that are used to train openAI models.

Then I define some project-specific knowledge, like the tech stack, and the project structure.

# Project structure
src
- app
- components
- lib

# Tech stack
- NextJS
- TailwindCSS
- Shadcn UI components

And finally, I added a descriptive text about what I wanted to project to be able to do.

Now, describing the project in such detail, with individual pieces of functionality, is not recommended in the long run.

I really only did this in an attempt to have it build the whole application from a single prompt. And it did that.

I would remove those specifics again if I want to continue development on the project.

# Project description
The app should output final format as VALID jsonl. Like the following example:
{"messages": [{role: "user", content: "user content"}, {role: "system", content: "system response"}]}

The web app will be a UI with input fields for messages to build out the jsonl.

- It should allow adding new examples for the AI into the jsonl. Which is done by adding a new line.
- It should allow adding more messages to an example by clicking a plus. Each message should then have a dropdown for whether the role is "user", "developer" or "assistant"

The prompt that built the whole app

Now, this is where I cheated a bit in the project description seen above.

When you clearly describe what you need upfront, you eliminate many of the issues that cause AI to generate unintended results.

This is pretty much all it took to build the functionality of the application:

And then a single follow-up prompt telling it, it made a small mistake.

I write a pretty technical message, but just an “you made an error” will usually do the job as well.

From JUST those two prompts, and in the span of 5 minutes I had this:

From there it took about 10 minutes of prompting to get it styled to the final version we see today:

So can you actually build a business utilizing AI?

Building with AI has become incredibly efficient, and I can’t wait to share more tips in the coming weeks.

The fact that you can build what I did, in just 30 minutes is amazing.

Is it a fully fledged SaaS business with business case and a paid userbase? No.

But is it a tool that people already want to use and pay for? Absolutely.

I can’t wait to work on many more SaaS tools, turn some of them into fully fledged businesses, and share the whole process right here.

See you next week 🚀