Game board is fed to ChatGPT with game element coordinates. GPT3.5 makes the decision, only keeping 1 message pair of history.
Most impressive is that GPT3.5 is outputting single word answers, which came from me adding the word "terse" to the system message.
My real goal is to use LLaVa to control the snake, but this was a neat first step to share :)