1. Data should be self-describing. No external tables should be needed to interpret the file (in contrast with GRIB).
2. Conventions should be developed only for known issues, and features should only be added as needed.
3. Conventions should be easy to use, both for data-writers and users of data.
4. The metadata and its meaning, as encoded in the conventions, should be readable by humans as well as easily used by programs.
5. Redundancy should be minimized to reduce the chance of inconsistency.

2007 Unidata NetCDF Workshop for Developers and Data Providers > CF Conventions