Does Vibe Coding Really Work? We Built a Game With Claude—Here’s How It Turned Out

There’s a brand new development amongst AI fanatics, and it’s not chatbots or brokers. We’re speaking about “vibe coding”—a session through which an individual merely talks to an AI, provides an thought of one thing they wish to construct, and begins iterating with the mannequin, asking it to repair and enhance issues alongside the way in which.

It is supposedly as simple as speaking to a good friend, and easy sufficient for even laymen with extraordinarily modest technical sophistication to do.

I wrote about it here. However to get past the hype, I wished to see if vibe coding may really produce one thing helpful.

Vibe coding a helpful app?

I selected Claude 3.7 Sonnet as my AI accomplice after Decrypt’s assessments confirmed it outperformed even Grok-3 for coding duties.

My software program mission was fairly easy: I wished to construct a typing sport through which phrases cascade down the display, tasking gamers with typing them earlier than they attain the underside to outlive.

We began with essentially the most fundamental immediate potential: “Write a ravishing, stress-free sport through which I have to kind out phrases quick to stay alive. The phrases are falling down and every letter has a vanishing impact as I kind on them.”

Inside minutes, Claude pumped out what regarded like a whole sport… However after I tried to run it, nothing occurred. The beginning button was lifeless. I discussed the error to Claude and it began to work assuring me it discovered the bug and was working to repair it.

Claude was executed, and the brand new sport was in entrance of our eyes. I attempted it, and nope… nonetheless damaged. Quite than diving into the code ourselves (that is not vibe coding!), I simply instructed Claude: “Nope. Nothing occurs after I click on the button. The sport by no means begins.”

The AI went again to work, and its second try really launched. I additionally cheated a bit and checked the code, noticing one other problem: The sport used solely about 10 completely different phrases, which might get repetitive quick. As an alternative of asking for a whole dictionary integration, I had Claude combine actual phrases with made-up ones—a sensible answer that stored the code easy.

I continued this back-and-forth with Claude, refining by pure language quite than code edits. Fourteen iterations later, I had one thing passable sufficient for me to share with out being ashamed.

When my colleagues examined the completed sport, the reactions had been principally constructive. One journalist even beat my excessive rating (which sucked), whereas others supplied concepts for enhancements. The one “bugs” they discovered had been really intentional limitations: randomly generated phrases, however no cell assist.

You possibly can strive the sport here. As of now, Ryan Gladwin holds the document with 2,830 factors, however hopefully I’ll declare my throne again anytime quickly.

Classes discovered

In spite of everything this and different experiments, listed below are some issues that may be helpful for these prepared to provide vibe coding a strive:

1. Fast prototyping works: We went from idea to playable sport shortly, with minimal technical enter. That is nice for novices with minimal coding talents. Nevertheless, it may most likely be an excellent method for skilled coders who want a tough draft to begin iterating.

2. Iteration is essential: The primary try hardly ever works completely. Success comes by conversational refinement and clear suggestions. On this stage, skilled coders may most likely go away vibe coding behind and follow regular debugging, utilizing AI correctly—a digital assistant that helps them with the method. Nevertheless, these prepared to commit absolutely to vibe coding might profit from utilizing very particular and detailed language. The richer and higher structured your immediate is, the higher the outcomes you’re going to get.

3. Do not rush it: Realizing when to simplify necessities helps obtain a working product sooner. Ideally, do not ask the mannequin to do every little thing directly. As an alternative, take into consideration the expertise as including completely different layers to a tough construction. Begin with the fundamentals, then ask the mannequin to repair one small bug, then one other, then add a particular factor you’d prefer to see—and transfer on from there.

4. High quality has limits: Whereas we created a practical and gratifying sport, it lacked the polish and optimization of professionally developed software program. Do not count on to develop Skyrim with Grok.

5. Debugging is conversational: Quite than looking for issues within the code, we merely described what wasn’t working and let the AI determine how you can repair it. This works properly, however after all, specialists might pace issues up by checking the code or speaking to the mannequin concerning the technicalities as a substitute of the anticipated outcomes.

Total, vibe coding can certainly produce practical software program if you happen to’re affected person sufficient to information the AI by a number of iterations. Is it production-quality code? Undoubtedly not. You will nonetheless want correct debugging, optimization, and varied specialised instruments to construct actually skilled purposes.

However then once more, the man who constructed Flappy Chook made tens of millions with a easy sport that did not want fancy structure or complicated programs. Typically, all you want is a good suggestion, and possibly—simply possibly—vibing with AI till you begin making tens of millions.