I Vibecoded a website. It only cost me my evenings, my sanity, and $120 of AI credits.

Let's be honest about what "vibecoding" actually is.

It is the collective delusion - shared by me, half of Reddit, and apparently a growing number of people who have never opened a terminal - that you can type a couple of prompts into an AI chat box and emerge, blinking, into the sunlight as the sole architect of the next Microsoft or Supergiant Games. Or at minimum, a moderately successful SaaS product that will let you retire to the Algarve by Q3.

Spoiler: you probably won’t retire to the Algarve by Q3. But you will build something genuinely impressive, and that is worth talking about.

Where I Started

My coding CV, before all this, was not exactly intimidating. A handful of mediocre WordPress sites. Enough CSS confidence to tweak a colour or nudge a border radius. A working understanding of responsive design - the kind where you remember phones exist before you publish, not after. Some vague awareness of how databases are structured. Enough Photoshop to produce graphics that wouldn't embarrass me at a bus stop.

HTML? Mostly avoided. PHP? Actively fled from.

That was the baseline. Keep that in mind.

The Setup

Early 2025. A few tentative experiments with AI chat models. A lot of YouTube videos - genuinely, YouTube is criminally underrated for this stuff. Eventually I landed on VSCode with GitHub Copilot, mainly because it offered pay-as-you-go rather than demanding a monthly subscription upfront before I'd proven to myself I wouldn't abandon the whole thing after a fortnight.

The interface looks extremely professional. Code pours into the centre panel - vast, scrolling columns of it, more than you could read in a lifetime. You start out thinking you'll review it. Check it over. Stay on top of things. You cannot, there is simply too much. What you can do is hit Ctrl-F, find a hex colour value, and change it to something slightly different. This feels, inexplicably, like a genuine contribution to the codebase. I recommend it.

What It's Actually Like

Your first prompt is ambitious. You know, "build me a cutting-edge web application that will take over the universe, and suck all the money out of the world economy into my back account" And the thing is - something appears, in a browser, that you can click on. After the initial "this is incredible, I have made a terrible mistake leaving my career so late" moment, you realise you need one more prompt. Then another. Then another.

You notice you've started phrasing things as questions rather than instructions. "Can you make it look good on mobile?" , "Can you add a login screen?" This is because you don't yet know what's possible, and you are, on some level, not entirely sure what you'd do if it said no.

Then you get granular. "Please can you move that label just slightly to the right?" And you notice you've started saying please. This is the moment you understand your true position in this relationship.

Four hours pass. A rate limit warning appears. You push through it. A second warning. You push through that too. By the third, the computer has decided that's quite enough for one day.

But the progress is real. Tangible. You can see the thing taking shape. You keep going.

Then GitHub sends you a notification that you've hit your $20 budget limit.

$20? You thought this was going to be pretty much free. You increase the budget limit and try not to think about it too hard.

Ten days later, you have a fully working application. It does what you wanted it to do. It looks, if you squint slightly, like something a professional made.

The Dilemmas Nobody Warns You About

What do you do while the AI is thinking?

A prompt can take anywhere from ten seconds to ten minutes to execute. The sensible answer is to go and do something useful - the hoovering, the dishes, a brief walk. The problem is that the AI will, at some point during your hoovering, ask permission to do something. Access a system file. Run a command. Delete a directory. That prompt will sit there, blinking patiently, until you return.

I once came back to a "Click to Allow" dialogue that had been waiting for twenty-three minutes. The AI had not given up. It was simply waiting. There is something quietly unnerving about that.

So you end up staying close. Watching. Which is, genuinely, largely wasted time - except for the moments it isn't, because sometimes the AI does something catastrophic to your codebase and you need to catch it before it compounds the error.

Back up constantly. I cannot stress this enough. Version control is your best friend and you should commit far more often than feels necessary.

The context window problem is one that catches you off guard. The longer a chat session runs, the more history the AI has to process with every new prompt. This slows things down and costs more. More critically, at some point the beginning of your session falls outside the context window entirely - and the AI forgets that you told it, in your very first message, never to use red. It uses red. You have been betrayed.

The solution is to work in manageable chunks. Newer models, such as Opus 4.6, have started doing this automatically, which helps enormously.

The money question arrives eventually. You've spent maybe $120 on AI credits. The project ran to a couple of weeks of roughly five-hour days. You've built something that, in a different era, would have required a development team of three people working for three months and billing you accordingly. Is your application worth more than $50? Probably. Does it make you more than $50? That is a question you should have asked at the beginning.

Oh, and it's addictive. I should have mentioned that earlier.

What I Actually Built

A blog and data platform - 4billionyearson.org - that fetches, analyses, and visualises data from sixteen different APIs across climate, AI, emissions, energy, and biotech. Built on Next.js with React 19 and TypeScript, deployed on Vercel, with Upstash Redis for caching and Sanity CMS for editorial content.

The AI chose the stack. When I queried this, it explained its reasoning in about four hundred words. I nodded along and trusted it, because what was the alternative?

Could I have built the blog part in WordPress? Yes, probably. Could I have coded sixteen API integrations, real-time data analysis, interactive maps, choropleth projections, and charted the lot? Absolutely not. Not in ten days. Not in ten months.

Does Anyone Actually Read It?

Yesterday, I posted one of the first blog posts from the site - written with AI assistance, obviously - into a Reddit group. Within 24 hours: 50,000 views, 110-odd upvotes, 50-plus comments, and around 300 people who clicked through and actually read the post on the site itself.

Is that a lot? In the grand scheme of internet traffic, probably not. In the context of a website that didn't exist a few weeks ago, built by someone whose previous technical achievement was a mediocre WordPress site - I'll take it.

It's a start.

The Verdict

Vibecoding is not magic. It will not make you a developer overnight, and you will spend meaningful portions of your time sitting anxiously in front of a progress bar, wondering whether the AI is about to delete something important.

But it is genuinely, surprisingly powerful. It is fast. It costs less than a decent night out. It produces things you could not have produced alone.

That last part matters more than the rest.

The author vibecoded this website. It is, in his personal opinion, pretty ok.