How to program it.
How to solve it. George Polya

Translating requirements into Demeter-F programs

Given an object o of class O, compute o.f().
Apply Polya's inventor paradox: Generalize the problem,
solve the general problem and map back to the
specific case.

Which information in O is really needed to 
implement f()? Which information is noise.
Which information in O is needed simultaneously?
At which node should the information be brought
together and computation be done on that information?

How many traversals? What are the subtraversals?

For each traversal:

Type unifying or type preserving?

Decompose into type unifying and type preserving traversals.

Type unifying (e.g., computing the average, container capacity checking)
What are the augmentors: which information is sent down?
What information is sent up? Use n-tuple, n>1,
if multiple
results are needed. How is the information combined at each node? 
Can you use a generic combiner? What is the fold, what is the 
default value?

Type preserving (e.g., copy an object with small changes, de Bruijn indices)
What are the augmentors: which information is sent down?
How is the information transformed?
Which default builders need to be overridden?