The biggest problem is that these programs replace the original version of TEMP. If you need it (such as if you kater discover a logic error), you no longer have the original input available.
Depending on the size of the data sets, you can be running up the bill. Each step needs to read the entire data set and output all the observations. It's easy to get into bad habits, such as using more steps that you actually need. Typically, input and output are much more costly than data manipulation statements.