Steal morphemes. Simple as that. Find a language, learn it's morphology and fucking steal away.
You'd have to come up with an algorithm to convert those morphemes to your phonetic/phonotactical framework. Sure it will result in something that would look unnatural (as in, such changes just don't happen in natural languages) under a close scrutiny by an experienced linguist, but it also gives you ways to add flavourful details on a whim, stuff like fancy alternations, semi-regularity, inject stuff that superficially looks like unproductive derivation, whatever.
To provide a quick example. Say, you have an isolating SOV language.
Inventory is m p b f n t d s r l k g h j w a e i o u
Phonotactics is C1-V-C2, where C2 is restricted to nasals and liquids.
Now, we decide to steal from Russian. The root "to sit" has at least 9 different forms (caused by palatalisations, alternations and other crazy IE stuff): sad, ses, sazh, sid, se, sizh, sed, sedj, sjad.
To properly convert these, we would have to do something with the impermissible codas. I suggest that we incorporate the following thematic vowel, if it's a verb, and one is indeed present. Otherwise, we use u as a velarised epenthetic vowel, and i as a palatalised.
Converting: sadi, sesi (the full word is sestj, the "tj" part triggers regressive assimilation), sagi, sidu, se, sigi, sedu, sedi, siadu.
There you go, a wonderful array of alternating roots that are going to make anyone question your sanity.
Also, I probably have a slightly unfair advantage of speaking four rather dissimilar languages.