You might find Ed Kirwan's thoughts illuminating:
   http://www.edkirwan.kgbinternet.com/fracdoc/frac-page10.html

One thought of my own that I haven't seen stated elsewhere in the same form
(though, of course, it may be common in material I haven't read), is that the
primary aspect of package design is the distinction between what I call
"published" and non-published items. A published item (method, class,
interface, ...) is something that you have advertised to client code, and which
you (therefore) have committed yourself to maintaining in essentially the same
form in all future revisions of the software.  "Unpublished" stuff is anything
else.  In Java that distinction maps fairly well onto public and protected
access for published classes and members, and private or "package private" for
unpublished ones.

The mapping is a bit loose, because:

   (a) a public member of an non-published class is not published itself
       (obviously).

   (b) sometimes it is necessary to make things public for technical reasons,
       even though you have no intention of publishing them (in my sense).

   (c) sometimes it is necessary to deprecate, and later withdraw, published
       items.

   (d) Java has no way of saying "this entire package should be treated as
       private".

(b) and (d) can only be handled (in Java) by documentation (as for instance in
the JRE where there are a lot of public classes which Sun regard as private and
have explicitly not committed themselves to maintaining over time.

With those niggles understood, one way of thinking about package design is that
it is all about deciding what you (as the designer of one chunk of software)
are willing to commit to supporting regardless (almost) of any changes to the
system (internally or externally to that code).

Packages are units of publication.  When you establish a package boundary, you
are signing a binding contract with the future.

I have never read the second of those, but I have both editions of the first
(and intend to get the third edition too) -- it is /excellent/.

   -- chris