Vibe Coding: Launching a Productivity SaaS in Just Three Days

8 min readAug 3, 2025

After my last article about the irreplaceable qualities unique to humans, today I’d like to share a hands-on experience that demonstrates just how far AI has come. This journey gave me invaluable insights as both a designer and a maker — insights I’ll expand upon in future posts.

TL;DR — Jump to the End?

I’m a product designer with a bit of coding knowledge, but far from a real developer. Until last Friday, I had never tried vibe coding. On Friday, I installed Cursor and took my first steps; by Sunday, my first functional SaaS service was complete — all without touching a design tool.

Here was my weekend:

Friday: Learned the basics of vibe coding, wrote out my PRD, and created a first prototype using Google AI Studio.

Saturday: Installed Cursor, connected Supabase for database, integrated Google Login, connected Gemini API, and set up Vercel for deployment.

I quickly hit the trial limits for Cursor Pro, so I paid $200 to upgrade.
Vercel deployment also hit its quota, so I paid $20 for more.

Sunday: Polished details and added extra features like an onboarding page.

When my laptop kept freezing due to Cursor’s intense background features, I impulsively bought a new MacBook (!)

The service is called LinkStash — a personal, intelligent repository.
Enter an article URL, and AI automatically assigns categories and tags. You can filter, track what you’ve saved or read with a calendar view, and get daily summaries. All data is stored per user.

Try it yourself: https://linkstash.vercel.app/

Just log in with Google (all we take is email and name, and data will be deleted if the service shuts down). For now, please use it on desktop!

How Did a Non-Developer Launch a Working Service in 3 Days? Where Is AI Now?

What is Vibe Coding?

If you work in tech, you’ve probably heard the term “vibe coding.” To me, it was like “bitcoin” a few years back — everyone talks about it, few have real experience.

Vibe coding means coding intuitively, driven by your ideas and intent, with the help of AI.
Instead of typing code line by line, you use natural language (Korean, English) to tell AI, “make a feature like this,” or “change the color to blue.”

You can create real, functioning results — without knowing how to code.
Of course, for highly robust, scalable products, manual coding and maintenance are still necessary, but the bar for product creation has never been lower.

I’d heard a lot, but never tried it myself. So last weekend, I decided to spend my unused vacation days building a product, just for fun.

Rapid Learning and Prototyping

Having bookmarked several posts on real-world vibe coding, I started there. Thanks to countless people who have shared their failures and tips, I avoided many early pitfalls.
As I learned, I used Dia, the AI browser — I could ask about any term instantly, speeding up my process.

Writing a PRD for AI

“PRD” = Product Requirements Document. Standard fare for anyone building features or products: goals, features, user flows, constraints, and so on.

One thing I learned from others: for effective vibe coding, write a PRD and design your DB schema in a way that AI will easily understand. I used a PRD Maker I found in someone’s article to generate prompts.

Google AI Studio for the First Prototype

I pasted my PRD into Google AI Studio’s “Build” tool, asked it to create the main page, and — presto — a working preview appeared.
Being able to skip straight from text to working software was mind-blowing.

Sampling “Lovable”

I signed up briefly for Lovable, another tool, but didn’t use it much (the free plan is very limited). Its design quality was just okay.
They say it’s better if you upload your own images and prompts, but I didn’t get that far.

Installing Cursor

Back to actual development — I downloaded the PRD and loaded it into Cursor, an AI-powered IDE for developers.

The editor looked intimidating — black background and white text everywhere — but really, all I wrote was a few sentences.
That’s the magic of vibe coding.

At first, it generated a page that looked like Lovable’s output.

Then iterated: whenever I wanted something changed or improved, I simply described it in plain language and let AI revise the code — over and over, until the result matched my intent.

Integrating Login, Database, and Gemini API

With help from Cursor’s chat and GPT, I integrated three things: login, database, and Gemini API.
Many of these services were new to me, so it took time to find the right API and secret keys.

But this mattered — linking login and the database meant every user’s data would persist.
I also paid close attention to DB schema design, making sure to split tables for each kind of data.

The Gemini API lets AI read any URL and auto-generate categories and tags.

Upgrading the Style

Once links were stored and filterable, I knew it was still pretty ugly.

So, I asked AI to switch to a Linear-style layout, making it a bit more minimal… but I still wasn’t satisfied.

I took screenshots of Apple UIs, gave those as examples, and told AI to implement a dark theme with glassmorphism. Normally, for a designer, changing all components and themes is a huge effort, but AI made these changes swiftly. With the right instructions, it cloned — or creatively transformed — design systems from GitHub with ease.

Infinite Visual Tweaks

The next phase was all about detail — making countless small design tweaks.
I could specify every visual nuance: component names, ratios, radii, button styles… all in plain English (or Korean!).

Sometimes, the instructions missed the mark, so I had to repeat or rephrase things — sometimes frustrating, but part of the process.

If I’d done it all in Figma, it might actually be faster. But there’s a catch: Figma doesn’t produce living components. Here, my goal was to avoid relying on a design tool, maximizing the power of AI, and connecting to open-source CSS/design system libraries whenever possible.

Subscribing to Cursor Ultra and Connecting Deployment

I used Cursor so much I immediately hit the free tier limit — so I upgraded to Ultra for $200/month.
That’s expensive, but if you think of it as hiring a developer, it’s a bargain. For someone who can’t code, this opens new worlds.

Deployment was another step — hosting locally was tedious, so I hooked up Vercel and, with AI’s guidance, deployed instantly.
Deploying so frequently, I reached Vercel’s limit and subscribed for another $20.

Polish and Feature Additions

Now it was time for those endless tweaks and micro-launches.
Sometimes, AI refused certain requests, or a change in one page impacted another — so instructions needed to be as clear as possible.

I even set a rate limit — five auto-categorizations per user per day — to avoid runaway cost if many people signed up and hammered the AI.
No critical bugs slipped through — AI handled things reliably.

The RAM Ceiling

Then my MacBook Pro started freezing — something that’s never happened, even with video editing. It wasn’t a storage issue (external SSD didn’t help either). It seems Cursor Ultra’s background agents ate up all the RAM. I wanted to disable them, but they kept running anyway — forcing me to reboot after every crash.
Eventually, enough was enough: I bought a brand-new MacBook!

Real Onboarding, Not Just Images

One thing I loved: I could use real, working components for the onboarding page. Onboarding guides new users to a service’s core features. Embedding live, interactive UI, instead of static images, meant I could test flows for real.