I recently got into vibecoding because i wanted to get an accurate score distribution for my favorite poker variant.
It's actually a really nice game once you understand the rules so i wanted more people to be able to try it. I recently expanding the evaluator to cover all variants of it and so far i have not found it to make a single error.
It actually started with me simply wanting to get an accurate distribution over the scores and i did get what's hopefully going to be an accurate distribution via chatGPT and claude (i asked bot to make a sim so i could compare results and hand evaluations). The distribution i got ended up being nicer than i expected it to be.
Later i made a version where you play against 3 bots but that did require me to provide a lot of guidance due to LLMs having a poor understanding of poker theory. I did however not see any need to make the bots super-good since the goal was to let people easily try out the game and for that purpose they are already good enough even if they are fairly easy to get an edge against.
https://nlp8.vintologi.eu/
Main thing that remains is to get a working sim for trying out the profitability of various starting hands (when the mediocre bots play them) but i am already quite satisfied with what i have been able to achieve just using free AI (i did however have to jump a bit between models due to free plans running out and due to some models not being able to do what i asked of them). I think i will need to redo the post-flop model from stratch since the model made by the AI was quite bad (that became very clear when i looked at the actual code).
People like to hate on AI but there is no way i would have been able to make this without it without spending weeks/months learning programming (especially not considering i had to both make C++ and HTML/javastript versions.
I did do some manual editing of the code on my own (such as to put links where i wanted them and to adjust MDF weights for the defense against all-in shoves).
One thing that did help was to have a detailed PDF explaining the poker variant (including examples) since that helped the AI models understand what i meant and also gave it examples to test it against. Still it did get some stuff wrong which i ended up having to ask it to correct later.
It's actually a really nice game once you understand the rules so i wanted more people to be able to try it. I recently expanding the evaluator to cover all variants of it and so far i have not found it to make a single error.
It actually started with me simply wanting to get an accurate distribution over the scores and i did get what's hopefully going to be an accurate distribution via chatGPT and claude (i asked bot to make a sim so i could compare results and hand evaluations). The distribution i got ended up being nicer than i expected it to be.
Later i made a version where you play against 3 bots but that did require me to provide a lot of guidance due to LLMs having a poor understanding of poker theory. I did however not see any need to make the bots super-good since the goal was to let people easily try out the game and for that purpose they are already good enough even if they are fairly easy to get an edge against.
https://nlp8.vintologi.eu/
Main thing that remains is to get a working sim for trying out the profitability of various starting hands (when the mediocre bots play them) but i am already quite satisfied with what i have been able to achieve just using free AI (i did however have to jump a bit between models due to free plans running out and due to some models not being able to do what i asked of them). I think i will need to redo the post-flop model from stratch since the model made by the AI was quite bad (that became very clear when i looked at the actual code).
People like to hate on AI but there is no way i would have been able to make this without it without spending weeks/months learning programming (especially not considering i had to both make C++ and HTML/javastript versions.
I did do some manual editing of the code on my own (such as to put links where i wanted them and to adjust MDF weights for the defense against all-in shoves).
One thing that did help was to have a detailed PDF explaining the poker variant (including examples) since that helped the AI models understand what i meant and also gave it examples to test it against. Still it did get some stuff wrong which i ended up having to ask it to correct later.