Does it make any sense to incorporate several approaches, and make the system adaptive?

Let's say we've done some benchmarking (which I haven't), and I have these systems, breakpoints and savings (negative is space saved):

string 
length  system
0-10    plain (+10%)
11-50   7bit packing (-10%)
51-100  huffman (-20%)
101-    LZ (-35%)
[download]

Now, the first 2 bits (or byte, if plain) indicates which system is being used. Note that the plain text system suffers because of the overhead of the leading indicator.

Would a system like this work? What would the breakpoints be? Is it worthwhile to combine any elements together?

Or is it the case that benchmarking should be used to find the system that gives the best performance on the typical data, and stick with that?