Post

Back on the AI wagon (again)

I've dabbled back into AI to help write code, again, this time it's apps.

Back on the AI wagon (again)

I’d gone quiet on the coding and AI stuff.

Not intentionally — just distracted, busy, and probably a bit overwhelmed. I’d been obsessing over ChatGPT and GitHub Copilot for a while and hadn’t really looked at all the other tools that keep popping up. There are loads of them now. It all felt… confusing.

Then, at a recent works do, my interest was piqued again by a colleague who’d casually mentioned they’d “made an app” using Lovable.

That stuck with me.

Prodding Lovable on the train home

On the train back from London I decided to give Lovable a go. I asked it to make a Blood Bowl roster creator.

The prompt was (almost intentionally) crap.

I wasn’t trying to be precise — I was just prodding to see what would happen with a really open brief. I didn’t feed it any data. I didn’t explain the rules properly. I just… asked.

And honestly? It was impressive.

In seconds — seconds — it produced something passable: https://lovable.dev/projects/8bb887f1-2a82-46a5-9e58-bae676312fcb

It wasn’t correct, or complete, or even especially good — but it existed. That alone was enough to hook me again.

Falling into Claude

I did some light research after that, and Claude was recommended.
By ChatGPT, amusingly enough.

I really liked the Claude interface straight away. It’s clean, calm, and easy to work with. I felt like I was making rapid progress almost immediately.

Before long, I’d got to this: https://claude.ai/public/artifacts/79290956-b0b0-43c7-a840-6efbde42a1f1

Claude helped me scrape data from sites, mess around with layouts, even try background images. The flow was good. Iterative. Encouraging.

But I hit a limit.

Publishing.

Claude was essentially producing one big JSX file. Fine for experimentation, but not great if you want something more robust — separate data files, images, features you can grow over time.

Taking the plunge with Cursor

That’s when I decided to take the plunge with Cursor.

You download it and suddenly you’re in something that feels a lot more like VS Code. Immediately more technical. More intimidating. More real.

The goal here was to build something a bit more solid:

  • separate data files
  • images
  • room to add features
  • something that could actually be published properly

Cursor helped a lot. It even rattled through setting up GitHub Pages publishing, so the app actually exists on the internet:

https://twoshedzz.github.io/blood-bowl-roster/

I hesitate to call it my app.

Is it my app?

I’ve vibe coded the whole thing. I haven’t written a single line of code myself. I’ve run a handful of terminal commands and that’s about it.

And yet… there it is.

Things I think I’ve learned (or am learning)

A few reflections so far:

  • Jumping straight in works… at first
    I bundled in and started describing the roster builder. That got things moving quickly, and that was fine early on.

  • Small, well-defined features work better
    The best progress came when I started asking for very specific additions, one at a time.

  • You need to know when a feature is “done”
    I’ve lost features a couple of times when adding new ones. In Cursor I even asked it to create and run tests. At one point I asked it to refactor the code so teams lived in separate data files — it did this really well.
    Two or three features later… everything was back in one file.
    I don’t fully know how. It could be me getting confused with the Cursor interface and branches, but it’s something I’m trying to keep an eye on.

  • You have to pay
    AI is addictive. Honestly, very addictive.
    It feels like being a product manager or MD running a tiny company with devs beavering away alongside you. Making something usable in minutes or seconds feels powerful — and fun.
    But you hit cost ceilings and token limits all the time.

What next?

There has to be a point now where I stop and actually review the code that’s been produced, and see what I understand myself.

I’d also really like to watch someone using Cursor who’s more experienced at coding and vibe coding. I’ve mostly stumbled around on my own, asking AI to interpret screenshots when I get stuck.

Part of me misses the days of sitting in an office with a team — being able to lean over and ask a colleague a question and learn something quickly, in real time.

This feels powerful. It also feels lonely.

More ramblings to come, I suspect.

This post is licensed under CC BY 4.0 by the author.