I've always wanted to make a game.

I've been experimenting with AI for the last three years. In the last few months, something shifted. It finally felt like there was an opportunity to really make something. Not a demo or a prototype, but a production-level experience that someone could actually sit down and play.

So I put it to the test. There was learning, experimentation, daisy-chaining tools together in ways I hadn't tried before. But it worked. And it makes me incredibly excited about what becomes possible when curiosity meets tools that are evolving this quickly.

For me, this is a story of empowerment. I've enjoyed painting throughout my life, and this felt very similar to that. Having your tools and making something you thought of, and at the end of it being able to reflect. It was a unique experience that I've continued to bring to other projects, but this was the first at this scale.

I made it dark on purpose.

Alan Wake was a huge inspiration. Light as the core of combat, not just atmosphere. I'd always wanted to build something around that idea. When I started this project, I also had a bias about AI-generated visuals: across 15 chapters, I wasn't confident the tools would keep hundreds of assets feeling like they belonged in the same world. Those two instincts pointed in the same direction. If the player can only see a small cone of torchlight, you don't need a consistent world. You need atmosphere.

The torch reveals hidden text on walls, solves mirror puzzles, pushes back enemies, and controls what the player knows about each room. Every mechanic connects back to that single cone of light. What started as both a love letter and a practical workaround became the thing that makes the game feel like itself.

Process, not prompts.

The prompts weren't the interesting part. The structure around them was. I ran it like a real project: clear phases, acceptance criteria, wrap-up routines that logged bugs. Every session started where the last one left off, not from scratch.

Four files ran the project.

No Jira, no Trello. Just markdown files that both the human and the AI could read, reference, and build from.

What it actually looks like.

Not a single magic prompt. Hundreds of short, specific exchanges. Describe the problem, get a fix, test it, find the next problem. The loop was fast. The discipline was in testing everything between each step.

Hundreds of exchanges like this across six weeks.

What was easy. What wasn't.

The things that seemed like they'd be impossible were often the simplest part. The things that seemed trivial caused the most trouble. That inversion is probably the most useful thing to know about building with AI.

There were real bugs too. Shadows initially appeared in front of walls instead of behind them. A quirk in the build tools meant restarting a chapter silently corrupted the level. The maths behind the lighting made some rooms completely invisible. None of these were hard to fix once found. The pattern was always the same: AI builds it fast, you test it carefully, you catch what it missed.