A few days ago, I lit the fuse on a working prototype of a government service. No team, no procurement cycle, no waiting for approval. Just me, a few prompts, and a handful of AI tools. And honestly? Fireworks. (Though if you’re looking for what happens after the display ends and what to reflect on how we turn these sparks into sustainable, governed public services, then you might be interested in my follow up – Beyond the Vibes).
Vibe coding (or vibecoding) is an approach to producing software by using artificial intelligence (AI), where a person describes a problem in a few sentences as a prompt to a large language model (LLM) tuned for coding. The LLM generates software based on the description, shifting the programmer’s role from manual coding to guiding, testing, and refining the AI-generated source code.
Vibe coding, Wikipedia
I’m not new to what’s now being called vibe coding. Over the last year ChatGPT has helped me to bring a few random ideas to life1. Last weekend I thought I’d see what Codex CLI could do and I was again blown away. I mentioned this at work and in the conversation that followed we mused on whether some of the frustrations we’d been feeling could be shifted by trying the same thing there.
So I sat down with a laptop, some product instinct, and a handful of different AI tools. I wanted to see whether we could finally conjure the ‘fireworks’ we’d been waiting weeks to set off. I started with ChatGPT and the scale of the task was a bit intimidating. But then I remembered about Firebase and in minutes had something to show off. As I did, another colleague responded by asking if I’d seen Stitch, and another colleague said I should check out Jules.
And once I discovered Jules, that was when things got really interesting. Very quickly I had something live. Not a sketch or simulation, but something real. It’s up and running on Render (and I’d love to give you the link but I probably shouldn’t let it escape into the wild; at least not yet).
Obviously it’s just a prototype. But that also seems to do it a disservice. What is true is that it absolutely appears to do the job we had in mind. No engineers. No designers. Just me, some prompts and decisions, and it works, and it works in a way that will absolutely elicit the right sort of oohs and aahs.
I suppose I ought to make one small confession. I really shouldn’t have done any of this. Inside the department, everything except Copilot is blocked (and even then you only get Copilot on a Windows machine, not a Mac). Which means this burst of delivery joy has happened off network, off platform, and probably against better judgment. But that, too, is part of the problem. When the path of least resistance leads outside the system, it’s the system that needs fixing, not the people finding their way around it. Well, I would say that wouldn’t I?
Now, for our purposes as a team this exercise might be the perfect fireworks but more broadly for government, what are the repercussions?
I’m going to call it: Jules and Codex earn their hype.
Continue reading