Cloud Clicker: The Cloud's Coffee Break

Because not everything needs a business case

Every post in this series so far has had a reason. Cut the hosting bill. Practice AI-driven DevOps. Build the pipeline I tell clients to build. This one doesn’t.

I wanted to build something purely because it sounded fun, and because I had an AI agent fast enough to make “just for fun” a realistic weekend instead of a someday-maybe. So: an idle clicker game, themed on the exact thing this blog is about. If Richardson Cloud is where I write about the cloud, Cloud Clicker is where you can zone out and reminisce.

Meet Cloud Clicker

Cloud Clicker is an idle game that walks through the history of computing: tally marks, mainframes, the cloud, and eventually quantum and satellite datacenters, thirty eras in total. It looks and feels like an Azure-flavored universe on purpose (it’s a cloud game, after all), but every piece of art in it is original.

Play it at cloudclicker.richardsoncloud.com, or from the small launcher tucked in the corner of this site. It saves your progress right in your browser. No account, no login, nothing to remember.

How it got built

Claude Code CLI built the whole thing. Same practice as the rest of this series: I picked the concept and the tone, an agent wrote the game logic and the tests, and I reviewed before anything shipped. It lives in its own repo, deployed to Azure the same unglamorous way the rest of my sites are: push to a branch, GitHub Actions builds and ships it, no server to babysit.

The part that actually took real trial and error was the art. Claude Code CLI didn’t draw any of it itself; it called out to Codex CLI, OpenAI’s own coding-agent tool, which happens to have image generation built into it, to produce a few hundred pieces across every era, an icon and a background apiece. My first attempts just described the style in words and called it done. Two batches, same written brief, came back looking like they belonged to different games entirely. One was busy and oversaturated, the other looked thin and washed out.

What fixed it: showing the AI real reference images every time instead of just describing the style, plus calling out specifically what had just gone wrong (“no purple takeover,” “no filled-in shapes,” “no text”). Once actual pictures anchored each run instead of adjectives, the drift stopped and the set finally looked like one game.

Tucking the game into this blog was its own small project: a launcher button that only loads the game once you click it, so nobody who never opens it pays any load-time cost for something they didn’t ask for. Getting the two sites talking to each other was a CORS problem, and a genuinely easy one: allow the right origin, and it just worked.

Back from the break

I didn’t walk away from this with a lesson plan. I walked away with a working game, a launcher button that took less code than this post did to write, and a slightly better instinct for how to talk to an image model.

Every other post in this series is about a real deadline or a real cost. This one only happened because the typing got fast enough that “I just want to see this exist” became a weekend instead of a quarter. All said, I scratched an itch I had for an idle clicker and had about as much fun building the game as I did playing it.

What would you build if nobody asked you to justify it first?