[dave333]

Inspired by your example I just had a go at creating a mountain climbing game in similar fashion and I agree it does well on the concepts but there are lots of bugs in the details. For example it referred to certain divs by class and id but failed to assign class or id attributes to the respective divs. Maybe there are separate elements generating the response that do not always know exactly what the other has generated. Given the elementary nature of the bugs it should be easy to add a code review element that can find and fix these bugs before presenting the response to the user. When I pointed out a failing it regenerated a new different version of the code responsible but missed the cause of the bug.

[neeldhara]

Oh yay, thanks for trying it out! Do you have your project up somewhere? I'd love to take a look.

TIL that apparently you can simply use the generic follow up prompt “did you do what I asked?” and it will verify and course correct on its own, see for instance:

https://twitter.com/ericjang11/status/1639882111338573824

No idea how well this works in practice for things like JS games with visual elements, but seemed like an interesting approach!

[dave333]

Still too many bugs to let this out in the world. I've probably spent 2 hours on it so far and the progress is amazing, especially compared to total inaction on this since I originally had the idea for the game many years ago (I'm now retired after a 40+ year software career). As an aside, one simple lightweight technique for bug tracking I have found beats everything for me is simply writing pages of one-line bug descriptions in a notebook by hand. Then when I fix the bug I mark that line with a check mark, and then when I have tested the fix I mark it with T. One thing the AI got right immediately was playing a sound file for the avalanche effect. Screenshot

https://ibb.co/CKDWdrV