(full disclosure - I work at Anthropic)
This was an experiment to try to generate a medium sized (240 line) program without writing a single line of code myself, only with prompting.
I had the best luck starting simple and gradually adding features (including starting 2D). Claude was able to debug its code about 2/3 of the time when it created errors, and 1/3 of the time I had to open up the code and help it more explicitly.
I documented some of my prompts and the code here: https://github.com/eschluntz/pipes-screensaver/blob/main/REA...
Nitpick, but this is what I hate about AI today. I've sat there hitting the "retry" button for 30 minutes on the same 5-line snippet of Python code before; it's an awful user experience, if not slower than writing it by hand. It's really aggravating when people create grand demos of their model product, and then disclose it actually took 8 hours of hand-correction and elaborate model jailbreaks.
It all leads to me being so thoroughly unimpressed by AI programming demos. I almost wonder if the Windows Pipe Screensaver was faster to program in the 90s than it is today with AI assistance.
From doing a lot of coding with AI I've built up a pretty good intuition of what's faster to do by hand (changing some hyperparameters, or fixing a 1 line bug) and what's faster for the AI.
In this case, I started in 2D and added complexity bit by bit. Being able to say "Now make it 3D" and have that just mostly work in 10s was pretty wild.
