This closely mirrors my experience using ChatGPT. It gets about 80% to the correct solution on the first go around, but then can correct it to be 95 - 100% correct with the correct prompting.