Vibe Coding Isn't Bad, You Are

I cannot get through a day without reading that everything produced by vibe coding is awful. The code isn’t production ready, config values are hardcoded, secrets leak out on the client side and entire databases are exposed. It’s best to not even consider a multi-tenant SaaS solution.

And we’re in a polarised world – so everyone is either for or against vibe coding.

The vibers? It democratises development, anyone can put their ideas into practice, and it saves a tonne of time testing a concept before investing months and thousands of dollars?

The haters? The quality is poor, developers will lose their jobs, it’s just AI generated slop, it’s impossible to finish off an idea when things get complicated.

Is vibe coding actually awful, or are you just doing it wrong?

It’s A Tool, Not A Religion

Vibe coding isn’t a philosophical principle. It’s a set of tools. If you gave me bag of tools and some raw materials and told me to build a house I’d laugh and politely decline. That’s not my skillset.

If you gave me a construction crew, told me that they’ll do what I tell them but no more and asked me to build a house – I might be tempted to have a go.

I’d most likely not consider foundation depth sufficiently, miss a huge amount of bracing in the roof and put bathrooms all over the place even if the pipework made no sense. I’d have an idea of what I wanted but at the end of it the house would fall down or not even finish construction.

I shouldn’t kid myself about construction. I’ve got a vague idea what a lintel is, I’ve got a reasonable idea of how plumbing and electrical wiring works, but I have none of the skills to design, oversee or construct a house. For that reason, I also am woefully unqualified to oversee a team of people following my instructions.

This is vibe coding.

You’re given an incredibly complicated team of experts that will help you but the gotcha is they’ll only do what you want and nothing more.

My Buddy Claude

I started working with Claude last July. I signed up, started a subscription and shouted “create me a game in the style of Monkey Island but based around a cat that rides a crinkle cut chip” into the ether.

I first had a yellow HTML canvas with rectangle on it.

Then I had several failed Unity projects that wouldn’t build.

Finally, I created a console app that was only pretty awful.

I’d spent $200 on AI credits. Oy.

We’ve evolved since then. I started using Claude to work through architectural ideas. I’ve written previously about the five-day coding binge where I worked with Claude to create my own programming language, compiler, assembler and virtual machine from scratch with zero experience. But Claude wrote none of it.

Over the last three months things have changed, and Claude probably writes 75% of my code. It’s a combination of model improvements and how I manage Claude.

I no longer treat it as a font of all knowledge – but rather I tightly define what I want to be done, I focus on testing, I review as if a member of my team had done the work and I iterate.

The key to it all though is that I am tightly defining exactly what I want it to create, I can verify that it is correct and it goes through exactly the same process as code that a member of my team would write.

I can do this because I’ve been coding since I was seven and thirty-five years later, I have a fairly good idea how this is supposed to work. But I still couldn’t vibe code a Unity app as I have no experience in that sphere.

With any tool you need to know how to use it. Vibe coding tools are no different.

I’m sorry – but vibe coding isn’t the problem. The users are.

If You Can’t Code, You Can’t Vibe Code

I can code. I get phenomenal results out of AI coding tools now. This year they have finally reached the ‘code per hour’ that I can produce giving me the ability to oversee rather than code.

But if you can’t code, that is not going to be your experience.

Sure, you can create a quick proof of concept and see if an idea works. But it will be no better than my theoretically wonky house. It will fall down.

To use any vibe coding tool, you need to understand code, architecture, security, performance, deployments and monitoring. If you don’t know what you need to ask for the AI won’t magically think about it.

And if you’re a developer and you know all that? It still might not be for you. What feature are you building? What does a user need? What should it look like? What kind of test plan do you need to follow? How should you sequence your features to engage with users more quickly and what adds the most business value?

None of those were technical questions but they all need answering if you want to build an actual product.

Not All Is Lost Though

If you want to vibe code, then that's great. Use it to learn. Ask questions. Read widely - not just about code but about all the disciplines that make a product. Throw away what you created and start again. Use multiple tools and learn how to master them.

The future absolutely contains 'vibe coder' as a role - but it will be someone with sufficient inter-disciplinary experience to oversee the AI tools they're using. You don't need mile-deep knowledge in everything, but you do need mile-wide, inch-deep understanding across the board to even have a chance of asking the right questions.

It's not a magic box. It's a set of tools.

Master them and you can create something extraordinary.

Shout into the void and a yellow HTML canvas can be yours.