The statistician or applications specialist wishing to implement a new technique in software has a bountiful choice of potential aids. These include both statistical software (well over 1000 R packages in the main archives, e.g.) and a wide variety of other software for nearly any likely aspect of computing with data. This welcome embarrassment of choice does prompt one to ask if some general principles can help to distinguish the better from the good, and to guide one's own programming. In this talk, we seek such principles, starting from our goal of exploring new ways to understand data, balanced against our responsibility for producing trustworthy data analysis and software. In several respects, the results prove surprisingly specific.