You could ask the LLM to first work out the solution in pseudocode, then translate to Pascal (or whatever). That way the variables are known after the initial pseudocode pass.

Human programmers also did this more frequently in those days than probably is the case now.