My Last Week and a Half: A Vibe Coding Journey

From hearing about YC founders using AI to "vibe code" real web apps, to going all in as a solo "non-technical" builder—what actually happened when I finally tried AI coding.

There's a product in my head. A web-based product that I want to use, and I think you'll want to use it too.

But here's the thing about getting an app out of your head and onto the internet: it's always been long, expensive, and reserved for those "technical" people. When I left Sheertex and was free to do whatever I wanted for the first time in a long time, getting this app out of my head and into the internet became a compulsion.

About three months ago, when I asked ChatGPT the best way to get this app made, it had many suggestions, but the main one was to use an app builder. Sort of like Squarespace but for app building. Specifically Bubble was the most recommended. These app building sites didn't exist the last time I wanted to build something, but it made sense to me that this was the way of the future.

So I booted up Bubble—think of it like trying to build a house with only LEGO blocks. Sure, you can make something, but it's slow, frustrating, and you can't really make it beautiful or get it to do exactly what you want.

So I gave up. I told myself I'd find consulting customers, hire developers, and wait for them to build my vision. Possibly forever.

Two weeks ago, a friend casually mentioned they'd been reading comments on Bookface (the Y Combinator portal), and apparently YC founders had figured out how to make AI coding actually work. They were talking about "vibe coding"—something I'd heard about but dismissed as not being for serious projects. Turns out these founders were building real, funded startups this way, not just tech demos.

That tiny comment lit a fire. Last Tuesday night, after 13 years of avoiding code, I opened my terminal.

Everything I'm about to tell you has happened in the week and a half since. And honestly, this setup has only even been possible since May.

My Complicated History with the Black Screen

Let me be honest about something: the terminal—that black screen with white text that every programmer seems to love—was a slow hell for me. It's basically a way to talk directly to your computer using typed commands instead of clicking buttons. Think of it like texting your computer, except your computer is really literal and unforgiving about typos, and it doesn't understand English.

It wasn't terrifying, exactly. It was worse. It was tediously, painfully slow.

Thirteen years ago, I'd tried to learn. I took courses with names like "Ladies Learning Code" and "Hacker You"—programs that really mattered at the time and taught me enough to be dangerous. But they also made it clear this wasn't for me. Back then, everything was Ruby on Rails and Heroku. I'd sit in those classes, watching instructors effortlessly type mysterious commands, and I'd follow along. But the moment I tried to build something on my own, everything broke.

Missing files. Cryptic error messages. Something called "dependencies" that were apparently missing but I had no idea where to find them. Database configuration that might as well have been ancient hieroglyphics. I'd Google the errors and fall down rabbit holes that lasted hours, only to emerge more confused than when I started.

Eventually, I just... stopped. I never fell for the bullshit idea that I wasn't "technical" but I put my brainpower into company building instead of debugging my terrible code. I founded companies that wrote code, I understood our tech stack, could talk through technical architecture, but always stayed safely away from actually writing the code myself.

The Magic Happened So Quietly I Almost Missed It

When I opened that terminal again after over a decade, something felt different immediately.

I did what any reasonable person does in 2025: I asked ChatGPT for help. Not just "how do I code," but specifically: "Walk me through starting a new app. Tell me what technology stack I should use. Give me the exact commands to type in the terminal."

The stack it recommended was completely different from my Ruby on Rails days:

• Next.js for the app framework (apparently this is what everyone uses now)

• React for the interface (finally time to learn what I'd been managing)

• Supabase for the database (I'd never successfully configured a database before)

• Vercel for hosting (Heroku's cooler, faster cousin)

• GitHub for version control (at least this one I remembered)

Here's the thing that blew my mind: the commands just worked. The same process that used to leave me crying with frustration now took about 10 minutes and actually produced a running app.

My Tool Evolution: From Chaos to Clarity

Stage 1: ChatGPT + TextEdit + Terminal Chaos

I started with ChatGPT walking me through setup, then asking Claude what code to put in my files. I didn't even have a proper code editor installed, so I was literally copying and pasting code into TextEdit. Yes, the same app you use for grocery lists.

But it worked. My app was running in my browser, and I was hooked.

Stage 2: Visual Studio Code Discovery

I knew using TextEdit wasn't how I used to do this, so I asked ChatGPT what people use now. It suggested Visual Studio Code. I could see it had suggestions, but they were confusing—I didn't want to accept any of them. So I stuck to using it just as a place to paste code, but it was nice that it connected directly to GitHub.

But I was still juggling: ChatGPT for setup questions, Claude chat for code (starting new conversations every 10-20 minutes when it hit context limits), a separate terminal window, and trying to remember where I put everything.

I was so excited about what I could build, even with my shitty setup, that one night over a game of euchre I asked my friend what app she wanted me to build. Between cocktails, I pumped out "Yay"—her gorgeous supplement tracking app—while we joked that her husband, who runs a software consulting firm, might need a new line of work. We started referring to him as "big software".

Stage 3: The Cursor Browser Breakthrough

Over lunch a few days into my adventure, I was telling new friends at Sunday House (my other post-Sheertex project) that I was down an AI coding rabbit hole. They asked if I was using Cursor. I said no. But the comment stuck with me (should I be using Cursor??).

I looked it up and found out Cursor is Visual Studio Code, but supercharged—it has Claude built right in. At first I used Cursor in the browser because I didn't realize it was a fork of VS Code meant to be downloaded as an app.

Suddenly I didn't need to manage multiple chat windows or keep starting new conversations. It was incredible.

Stage 4: Cursor Desktop + Better but Still Chaotic

Then I discovered Cursor had a desktop app, and it was even better. It was automatically committing my changes to GitHub, instantly updating my live app on Vercel.

But Cursor told me my app had many issues that were clear symptoms of rapid development. I couldn't help but be frustrated with it telling me that—I mean, you built it!!!

I was still juggling: ChatGPT for setup questions, Claude chat for code (starting new conversations every 10-20 minutes when it hit context limits), and a separate terminal window for running commands.

Stage 5: Claude Code - The Game Changer

At this point I realized I should maybe do some research into how people were actually vibe coding. So instead of playing Grey's Anatomy in the background, I turned on YouTube tutorials.

People kept mentioning Claude Code in the tutorials, describing it as even better than Cursor. The setup seemed intimidating (you had to run it from the terminal), but I was watching a tutorial video and saw someone else's setup—that's when I realized I'd been making this harder than it needed to be: you can run the terminal right inside Cursor. Then create a new terminal window and run Claude Code in that, in Cursor.

Basically now I would be able to chat with Claude Code, in Cursor, and Cursor, which was actually chatting with Claude too. While running my terminal from the same window. Very meta. I started using Claude Code's chat instead of Cursor and stopped using Cursor's chat entirely.

Claude Code was a revelation. It seemed to understand my whole codebase better than Cursor. When it ran out of conversation space, instead of forgetting everything, it would intelligently compact our conversation and our progress and keep going.

I rebuilt my entire app using Claude Code, and for the first time, everything worked consistently.

Stage 6: My Current Setup (The Best of Both Worlds)

Here's where I hit my one major wall: styling. Claude Code is brilliant at logic and functionality, but ask it to make something visually stunning? The results look like they were designed for 1995. And Claude Code wasn’t committing its changes to git hub, which I thought it would be, so when things went bad it was nearly impossible to go back. In frustration I opened my Cursor chat again for the first time in a couple days—Cursor nailed it. And I have now started using Claude Code and Cursor chat, helping the other when one gets stuck.

So now I have a hybrid workflow:

• Cursor Desktop App: Code editor / interface for terminal

• Claude Code for backend logic, database setup, API connections

• Cursor's built-in chat for frontend styling and making things beautiful, helping when Claude gets stuck.

• ChatGPT For questions and guidance on setup / architecture

The Real Breakthrough: Function Over Form

For years since giving up on code, I could only build websites in editors like Squarespace, they could be pretty, but basically had no functionality. I would always have to go to a technical person for anything that actually worked. Now I can build account management systems, store data in databases, create API integrations—myself, FAST.

Things like Supabase for databases have always sort of existed, but are still complex to configure. Now I can just say what I want in English and the system configures itself. Can we talk about how insane this database situation is? Back in my Ruby on Rails days, database configuration was where dreams went to die. Hours of setup, mysterious configuration files, local vs. production environment nightmares.

Now? I tell Claude what I want to store, it writes the database schema and connection code, and suddenly I have a fully functional database with a beautiful admin interface. I'm storing user data, handling authentication, managing complex relationships between data. Things that used to take me weeks of frustration happened in an afternoon.

Over that same euchre game, my friend asked what about security? I said—and I really do believe this—that the AI is going to do much better security than I ever could. I can actually analyze for security risks rather than asking Bob to give me his opinion. Her husband "big software" begrudgingly agreed.

The Real Cost of Vibe Coding in 2025

I've spent about $140 on Claude credits this week, plus monthly subscriptions to Claude Max ($75/month) and Cursor Ultra ($200/month). That's real money, but let's put it in perspective: three months ago, I was resigned to hiring developers and waiting indefinitely. Now I'm building exactly what I want, exactly how I want it.

My code is definitely messy. When "we" get things wrong I can tell sometimes old versions still hang around in the code. I can tell the whole codebase isn't getting analyzed every time, so sometimes things we already have get forgotten and rebuilt. But for the first time everything feels possible, I'm moving fast, and I'm actually having fun.

I will likely build it and rebuild it a couple times as I refine the actual architecture—something you used to have to plan beforehand because a rebuild was an insane and terribly long process, before these tools. But I feel like once I get to where I want to be architecture-wise, I will be able to produce a stable version.

Why This Time Is Different

If you tried to code before and ran away crying (literally, in my case), I need you to understand something: it's not just that the tools got better. The entire paradigm shifted.

"Vibe coding" isn't traditional programming. You're not memorizing syntax or debugging cryptic errors for hours. You're having conversations with AI about what you want to build, and it's translating your ideas into working code.

The barriers that stopped us—the cryptic error messages, the missing files, the feeling of drowning in complexity—they're still there, but now you have AI pair programmers who never get frustrated with your questions, never judge you for not understanding something "basic," and can explain anything in terms that actually make sense.

But here's the thing: much like great writers will be the best AI writers, great coders will be the best AI coders. You need to know what "good" looks like, or at least it's WAY better if you do. You have to know the questions to ask. The fact that I have any background makes me more powerful than if I didn't know the questions to ask, or didn't know the architecture I wanted. This is what humans need to get good at.

If someone who used TextEdit to write code files and had never successfully configured a database can make this work, so can you.

The future belongs to builders, and the barriers to building have never been lower.

It's time to stop waiting forever.

---

I'll keep sharing what I learn—the victories, the frustrations, the tools that actually work. This is all changing so fast, I'll keep you posted as I learn and as the paradigm shifts, but right now my goal is to produce a stable app I can share with people to test by the end of the month (1 week to go). Because the only thing better than finally building what you want is helping other people do the same.

Oh, and that product in my head? It's called Oomira, and I'm literally building it now—in another window, while also writing this article with Claude (ChatGPT struggles with long-form narratives and lacks personality—Kat is saying this, not Claude marketing itself lol). If you want to see what vibe coding can produce, join the waitlist at oomira.com (where you can enter yourself into my live working database :P) to get notified when it goes live.

---

P.S. I need to end this post with a rant because I think it's critical to the future of western world prosperity that I get through to you on this.

We need to stop calling most software companies "tech" companies and valuing them with tech premiums. The vast majority of what we call "tech" today are just web businesses. The only things that should be called tech and get tech multiples are what we currently label "deep tech"—advanced AI systems, quantum computing platforms, biotech tools, space exploration, advanced manufacturing. Ironically, these actually get lower multiples because they're hard and expensive to do—but that's how you know they're actually tech.

The valuation divide is insane: a regular business gets 10x profit, a tech company gets 20-500x revenue (yes, I literally heard of a company getting 500x this week). Not only should web businesses be valued like regular businesses, I'd argue most should be valued at less than the dry cleaner down the street because that's more defensible these days. My week-and-a-half coding journey proves this—if someone using TextEdit can build functional web applications this quickly, then building web apps has become commodity infrastructure.

A couple weeks ago I sat at a conference as 2 out of 3 VCs on a panel proudly said they don't invest in deep tech; the third barely did. Venture capital was supposed to fund new frontiers, but it's become about opportunistic financiers benefiting from inflated multiples driven by acquisition arbitrage and public markets not yet realizing (or not wanting everyone else to realize) how indefensible and short-term most web application life cycles are, not long-term value creation. Most web apps end up getting killed after acquisition anyway.

VCs need to stop paying tech premiums for web businesses. We need to stop rewarding non-tech companies with inflated exits. And we need to start investing in actual tech companies that will actually give the Western world a technology advantage and long-term prosperity. These are possibly software companies but more likely they are full-stack businesses. Everyone thinks I'm saying this because I have scars from building Sheertex—and it's true, but it's also how I know I'm right.

P.P.S. Just as I thought I was ready to push this post and my app live, Cursor made a mistake and we lost everything. I thought it had been committing as we went—it hadn't been. It reverted to a very old version of the app, and I had to spend an hour sleuthing, debugging, and almost crying to get it back. Note to self: commit to GitHub more often. Even in the age of vibe coding, some old habits still matter.