Minesweeper 95

I wanted to try the newly released Gemini 4 Argon in Google Antigravity, so I built something nostalgic in an evening: Minesweeper. A familiar game gave me a concrete starting point and an outcome I could try for myself.

The result has a Windows 95 look, Beginner, Intermediate, and Expert boards, dig and flag controls, and weekly and all-time leaderboards. I didn't spend much time writing the spec, and I'm happy with how it turned out.

The session still followed a structured process: specification and clarifications, an implementation plan, a task breakdown, then implementation, testing, and build. For this experiment, moving through those stages felt quick.

Spec through build

~20 minActive agent time

Implementation

15m 52sIncluding TDD and build

Input from cache

84.7%Across the full session

These figures come from the Antigravity session breakdown. The phases from spec through build add up to 20 minutes 16 seconds of active agent time, across 24 minutes 9 seconds elapsed. They exclude the later usage-analysis turn and do not measure my personal spec-writing time or the full time to deployment. The cache percentage covers the full recorded session, including usage analysis.

I think the result also benefited from the problem being familiar. Minesweeper has established rules, and it is easy to picture what the finished game should do. Much of the context was already available, so my spec could be short.

My takeaway is that AI tends to have an easier time producing a good result when the problem and expectations are clear. For a first experiment with Argon, I was pleased: relatively little time explaining what I wanted led to a game I'm happy to share.