Is it correct or incorrect to split data into train and test portions before using them on tumbling and sliding windows?

I can't find any book, article, or trustworthy webcontent on the correct way of handling timeseries data for inputting it on a LSTM predictor. I did write code that first splits the first 80% of the data for training, and 20% for testing, then create windows inside each portion, both tumbling windows and sliding windows, and make predictions. The results are not good and a PhD criticized what I did as wrong, saying that when data is windowed, it should not be split in 80%/20% before, because the windowing itself will handle this. But I can't find references to support what he said, so I cannot fix my work! I need to find the correct way, but I only have the wrong way in my hands. Help! Please, point me to something helpful!