Behind the Buzzword: What is vibe coding?

In this edition of Behind the Buzzword, we cover a controversial topic in the software development community: vibe coding. Here's what you need to know about it, explained in a four minute read.

What is vibe coding?

Vibe coding is a prompt-driven approach to software development. You tell an AI what you'd like developed, and the AI writes the code for you. Usually this is done in a few lines, or "vibes." For example, you might type in to ChatGPT:

 “Build a React component that displays recipes, including ingredients and instructions.”

… And the AI gives you the code for that. If it doesn’t run, you ask the AI to fix it until it does, following the "vibes".

The term was coined by Andrej Karpathy, former Director of AI at Tesla, who said vibe coding is more mood than actual know-how. Below is Andrej's original tweet:

Vibe coding VS traditional development VS AI-assisted development

Vibe coding is significantly different from traditional development or AI-assisted development:

• With traditional development, the developer writes all the code, and has full control over every detail. This is what most people think of when they think of coding.
• With AI-assisted development, the work is split between the developer and an AI assistant. The most common form of this is AI code completion, where a developer writes codes and an AI fills in the blanks. With AI-assisted development, the developer still validates each line the AI suggests, and sets the overall direction.     
• With vibe coding, the developer asks an AI for the finished product, and then trusts whatever the AI produces. Any errors are solved by asking the AI to debug the problem. The developer typically does not know what the code does or how it operates---only that it works. 

To coin an old saying, traditional coding is "measure twice, cut once." AI-assisted development is "I'll measure, then get an AI's help to cut it." Vibe coding is "I'll ask the AI to measure and cut it, while I drink some tea."

Is asking an AI to produce code and checking it the same as vibe coding?

No. In the first scenario, it's still AI-assisted development, because you're asking an AI for help but still validating the output (and typically editing the code.) With vibe coding, you wouldn't examine the output at all, because you're trying to forget the code even exists.

What are the benefits of vibe coding?

The potential benefits of vibe coding are rapid prototyping and iteration. It can be very fast to build something that runs, and that’s great for a proof of concept. Individuals with limited coding experience can theoretically contribute more directly and developers can focus on more strategic or creative tasks.

What are the pitfalls of vibe coding?

While it may seem like vibe coding saves time and makes things easier, there is evidence that suggests that it simply creates more work further down the line in the form of vibe debugging. On top of this, developers often underestimate the time they spend generating prompts and waiting for them to compile, which tends to result in slower development, not faster.

What is vibe debugging?

Vibe debugging is the term given to the often laborious task of debugging AI generated code. With traditional development, a developer knows what their code does and (ideally) leaves notes. But because AI does not "think," the code it produces can be convoluted, insecure, and error prone, requiring excessive rework.

Vibe debugging typically requires someone with traditional programming skills to demystify and fix the code, which may take more time than just writing the code from scratch to begin with.

To use our earlier analogy, while the vibe coder was drinking tea, the AI has build a very bad house, and now a proper builder has got to go fix it: and as anyone knows, it's easier, cheaper, and more efficient to build a house right the first time than to fix a badly built one. 

Why coding with AI can be slower, not faster

Research has found developers are prone to misjudge how fast an AI speeds up completing their tasks. In a recent study by METR, developers estimated a 20% increase in task completion speed with AI, while tasks were actually being completed 19% slower.

Why? Any time they saved in actively coding, testing, and researching was lost in reviewing AI outputs, prompting AI systems, and waiting for AI generations. Meanwhile, developers only accepted less than 44% of code generated by AI without modification, and the majority report making major changes to clean up AI code.

It's worth noting this was a report studying the benefits of using AI tools as a developer, not specifically vibe coding. However, it showcases a trend of developers overestimating the productivity benefits of using AI, while instead losing it in new prompt-specific tasks.

(Note: People have a LOT of misconceptions around their abilities when it comes to AI, and it isn't just limited to how fast it makes them, either. To learn more about that, read Pluralsight's AI Skills Report.)

How to implement prompt-driven development without vibe coding

• Actually try and understand what the AI is making, as you will likely need to debug it.
• Be clear and specific about what you want.
• Set context and constraints, like declaring the tech stack and sharing existing code.
• Ask for small components of your end product---don't ask it to make everything in one go.
• Iterate, test, and refine what you build.
• Independently measure your productivity gains rather than relying on self-assessment.
• Look out for any security gaps in the code, including hardcoded secrets and credentials.

Conclusion: Less sipping tea and vibes, more watching and understanding

AI does not think or reason---it's just using complex pattern recognition. As a developer, your knowledge and critical thinking skills are still required to validate inputs and outputs. While it might seem nice to just leave it up to an AI and follow the vibes, chances are your vibe will be ruined later on with excessive debugging---and you may not have saved as much time as you think.