I think good problems should be as short as possible, like Fermat's last theorem or the Goldbach conjecture. Don't tell stories, nobody cares about the fxxking stories. Just be straight.

Besides, they should contain some novel ideas and/or novel applications of existing theories.