It feels like something like this can be hacked together to be more reliable with some image to text generation plugged into the existing ChatGPT, and enough iterations to make it robust for these how-to applications. Less Turing-y but a different route to the same solution.