Word Construction for a New Conlang

I'm finding myself in a state of throwing away all the beginnings of
conlangs I've started and wanting to start over again.  And I'm just
not sure how to start, so I turn to you (pl) for help...

It's the phenomenon of making words and making phonologies.  Think
back, if you can, to when you first started on your conlangs (or first
started on a new project)...  How did you go about making a phonology
and then making words?

Let's assume for the moment that the thing to do is (1) settle on a
phonology and phonotactics and (2) make words according to that
pattern.

Not being a professional linguist, I'm not sure I've ever seen a
complete description of the phonotactics of a natlang, so I'm not sure
how I would go about describing such a beast.  I don't explicitly know
the phonotactics of any of the few natlangs I've studied, nor of my
native language.  As for phonologies, I've seen phoneme inventories
but I'm not sure I've ever seen anything like a complete analysis of
the phonology of a natlang (allophones and all).

Given a "sound" that I might want a language to have, when I try,
clumsily, to analyze that sound and make a formula for creating it, I
usually don't end up getting anywhere near it.  And when I try to make
a system from scratch, not following a vision but just making up some
rules and seeing where they lead, I don't often get anywhere I want to
be.

I've tried using Jeffrey Henning's _Langmaker_ and Christopher
Pound's  _werd_ and you can do some cool stuff with them, but again, I
suffer from an inability to come up with formulas for word creation
that produce quite the sets of words I want.

Do the rest of you start with this kind of harcore linguistics
description and proceed to specific words, or vice versa, or somewhere
in between, or some cycle between the two, or something entirely
different?  Where did your words come from?

I'm curious how other people work, especially people who, like me, do
not have a really hardcore linguistics background with exposure to a
lot of natlangs and formal descriptions thereof.

Ed Heil ------------------------------- edheil@postmark.net
"Facts are meaningless! You can use facts to prove anything
   that's even _remotely_ true!"           -- Homer Simpson