Consider implementing a program to just copy data.

$ nccopy   input_file   output_file

1.     create empty output file
       add dimensions and global attributes in input to output
     
       for each variable in the input file
            add the variable definition to the output file
            copy the variable's attributes to output
            copy the variable's data to the output
     

2.     create empty output file
       add dimensions and global attributes to output
     
       for each variable in the input file
           add the variable definition to the output file
           copy the variable's attributes to output
     
       for each variable in the input file
           copy the variable's data to the output
     

If output is a classic netCDF file, which method is faster, and why? If the input has record variables, can the faster algorithm be improved?

For HDF5-based netCDF-4 files, schema changes such as adding new variables are efficient, because both metadata and data can be appended to a netCDF-4 file.

2012 Unidata NetCDF Workshop > Formats and Performance