Not looking at the code

Despite the town criers' warnings, vibe coding is here to stay. Worse still—it's effective! I have not yet built a ten-billion-dollar ARR SaaS product, nor have I successfully used it to manage a 10,000 microservice cluster of pure, single-line node functions. However, I have built a profitable software agency and I just shipped a fully-featured game (with tests!), so take my experience with a pound of salt.

Claude Code has created possibilities that previously just weren't possible. The human limit of having numbered hours in the day no longer apply. I can now pursue small, trivial projects that previously would've been very labor-intensive. So, I built botnet, a game loosely inspired by LEGO's Spybotics: The Nightfall Incident (the link is to a faithful remake on Itch.io from a fan).

While the gameplay of Spybotics is hardly a novel concept, its combination of music, unique artwork, and fun difficulty-scaling created a real nostalgia from my childhood that I wanted to capture and share with my son.

I started off simple: HTML, JS, and the Phaser.js game engine (which turned out to be an over-complication right off the rip, but I'll get to that later.) I had a working prototype within 5 minutes, which quickly bloated to an unmanageable mess after 2 weeks of vibing away.

One day, I realized... I never even looked at the code.

What shape was it? Could I read it?

Turns out, not only did I not care, it didn't matter. Mistakes don't really matter because you can clean them just as easily as you create them. Notice a pattern that keeps tripping Claude up? Ask it to identify the pattern, make a note to avoid it, and move on. Just like normal human coders.

When vibe coding, the harness matters a lot. The "harness" is the combination of all the ways the model can check itself and its code.¹ The problem with the Phaser.js game engine was that the entire game existed in an HTML canvas that Claude had no way to debug or check for correctness. So I buttoned up my YOLO trousers and hatched a new idea: refactor the entire game from app.js and Phaser to a React-based renderer with a pure JS game engine. Roughly ~6 hours of wall-clock time later, I had 500+ tests verifying game engine functionality, and verifying the React rendering system.

Oh... and I also built an entire new roguelite game-mode which eventually replaced the "classic" game mode.

I've played an additional 2-4 hours of the game since then and haven't noticed a single true "bug". Here and there I've tweaked UX, and I still need to do another pass on the difficulty scaling, but the core game works and is stable.

I consider myself lucky that I was never a great engineer. I can't recite algorithms very well, nor am I particularly good at building.

But, I do love building. And it's never been easier (or more fun) to build.