What building with Opus taught me about AI intuition

A week ago, HexaScaler was just an idea, a 3D hex-tile strategy game about building and running your own data centre — power, cooling, GPU capacity, compute economics, the works. Today, it’s a genuinely playable 45-minute experience that levelled up my experience of vibecoding.

Built with Claude Opus 5 in Claude Code, the process changed how I think about what “collaborating” with an AI model actually means. Opus 5’s intelligence and uncanny ability to understand where I was headed with my thinking around a complex task was exciting, if slightly eerie.

Thinking alike

The first thing that struck me was intuition. I didn’t have to explain what a hex-tile strategy game should feel like with all the adjacency rules, the resource loops, the tension between expansion and upkeep that makes the genre work. Opus 5 already understood the conventions of hex tile games (hexagons are bestigons after all), and reached for them naturally, the way a designer who’s played a hundred of these games would. I spent far less time specifying mechanics and far more time steering direction and polishing the experience early on.

The second thing was how it handled vague feedback, vague in the sense that sometimes specific changes were difficult to describe even with screenshots. Game feel is notoriously hard to articulate, “this turn drags,” “the power system doesn’t feel tense enough,” “I want it to feel more like Power Grid here.” These are the kinds of notes that usually take several rounds to land. Opus 5 consistently got the intent right on the first or second pass, translating fuzzy, half-formed impressions into concrete changes to pacing, numbers, or UI.

Getting to know each other

And critically, the game’s balance defines success or failure in real world play. Claude ran hundreds of simulations itself, but a real user experience is what matters, picking up on nuances code just can’t. Yet that too was a pleasant experience and the collaboration and learning on both sides gave me the strongest indicator yet of near harmonious agentic working.

What that adds up to is speed without sacrificing good judgement. From in-the-brain concept to a working, actually fun, 45-minute play session in a handful of days, not weeks. Not a prototype held together with placeholders — a session I could hand to someone else and watch them play through to a real ending, and learn about data centres too.

I’ve used a lot of AI tools for a lot of tasks. This is the first time one felt less like a very capable assistant and more like a collaborator with taste.