Since my last post, I’ve been doing more experimenting with both Claude Code and GitHub Copilot CLI. I’d been having trouble coming up with any good ideas for little projects to work on, but I recently watched an episode of Astrid, in which the main character is working on a Towers of Hanoi puzzle. I’m pretty sure I had a little BASIC program for a Towers of Hanoi game running on my TRS-80, a very long time ago. I hadn’t thought about it much since then though. But it seemed like a good little old-school challenge I could use to spend some more time with AI coding tools.
I started out with Claude Code. My methodology was to start out by writing a “requirements.md” file and placing it in a new folder. I then launched Claude Code in that folder and asked it to generate a “PLAN.md” file, with a detailed plan to create the application. I then iterated through the steps in the plan, letting Claude do most of the work. That worked really well, and I had a running, bug-free, application in just about an hour. Checking my usage, I saw that I’d used just a third of my session limit and 4% of my weekly limit. And I was using whatever the default model was, probably Opus 4.6? You can see the whole project here, if you want.
After that, I tried the same experiment with Copilot CLI. That didn’t go nearly as well. The generated PLAN.md file wasn’t bad, but it was a good bit less detailed than Claude’s. And, while I got to a “finished” application pretty quickly, it had a fair number of bugs in it. I spent a little time working through those bugs today, and I have a finished application now, though it’s still a little rough around the edges. In terms of usage, I’m at 50% of my monthly max (though of course I’m on the free plan here). For the model, I used the default, which was “claude-haiku-4.5”. I think that may account for a good bit of the difference I saw between Claude and Copilot here, though it might not be all of it. I think that, if I paid for a Pro plan and switched to Opus 4.6 here, I would have gotten results similar to Claude. I’m not sure though.
So where to go from here? In both of these cases, I had asked the AI to create a C# console app for the game. I’m considering going back to the Claude one, and asking it to rewrite the user interface using a TUI library, like Terminal.Gui or Spectre.Console or maybe Consolonia. I haven’t played with any of these, but I’ve been curious about them, so that might be interesting.
I’ve also considered asking Claude to add a SPA project to the solution with a web-based user interface. I could either tell it to use a specific technology for that, or leave it open and see what it does. That might be fun.
I could also bite the bullet and pay for the Copilot Pro plan, and then repeat the exercise from scratch in Copilot with a better model and see how that works. And/or take the PLAN.md file from the Claude implementation and let Copilot use it. It would be interesting to see how much of a difference a good, detailed, plan makes.
Well, I’m off today for Presidents Day, and I’ve already spent a good bit of the morning messing around with this stuff. Maybe it’s a good time to have lunch, and then switch to just relaxing for the rest of the day.
