Things change fast in artificial intelligence. Around two years ago I wrote an article here on the limitations of using AI to produce computer code. I concluded that AI would work best as a tool to help coders in their work, rather than as a replacement for humans. But the world has moved on.
Vibe Coding
A year later, ‘vibe coding’ was the term on everyone in the industry’s lips. In an X post in February 2025, OpenAI co-founder and former Tesla AI director Andrej Karpathy said ‘There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists… I’m building a project or webapp, but it’s not really coding – I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.’
Turning Karpathy’s comments into English, the idea is to give a detailed prompt in plain English of what you want an app to do to an AI, then don’t look at the code it produced, but rather refine the result using follow-up prompts and noting things that went wrong until you get a usable app.
Journalist Thomas Claburn who writes for The Register tried vibe coding and referred to AI’s ‘indifference to craft’, concluding that ‘Working with an AI model like Claude Code is difficult because you have to hold two contradictory thoughts in your head – the model is both highly capable and utterly clueless.’ It could both make errors no human programmer would fall for, but would also add aspects to the final app that he had not considered.
A Useful Way to Build Small, Quick and Dirty Apps
Inevitably, professional programmers criticise the output of vibe coding – with good reason. This is a useful way to build small, quick and dirty apps that aren’t mission critical to an organisation. It’s arguably closer to the way we used to automate activities in programs using macros than serious code writing – but taken very much to the next level. As such, there is no doubt that some tech-savvy non-programmers can use vibe coding to put together tools that enhance their own way of working.
The best analogy to vibe coding is probably the use of 3D printers in companies that produce physical goods. They are very useful in research and development, when playing around with new ideas on a small scale. But in most cases, it would be impractical to use 3D printing to produce a finished product, as it simply doesn’t have the accuracy or complexity to, say, build a complete washing machine or car.
“The Vibe Coding Hangover is Upon Us”
Similarly, professional coders may use vibe coding to play around with new ideas. But it is not an approach that can be used to produce robust, sophisticated programs. And because the person building the app has no connection with the code, it is not practical to take this approach when it comes to maintaining and enhancing a product. A 2025 Fast Company report The Vibe Coding Hangover is Upon Us quoted a senior software engineer saying ‘Code created by AI coding agents can become development hell’: the approach makes maintenance of software more labour intensive. AI-produced code also tends to have significantly more security issues than the human generated equivalent.
Overall, Karpathy’s original assessment that vibe coding was ‘not too bad for throwaway weekend projects’ remains a good assessment. No one is going to vibe code the next spreadsheet, booking system or large language model. And there remains the challenge of keeping AI-generated code working. For now, at least, we still need the human hand on the keyboard.