hehe, yea, that would help to explain that part.
What I do is parent a locator to the original camera and translate it around .01 in front of the camera. Then look through the smoothed out camera. Now use the "shakeExport.mel" that is here on highend. Use it to export the x,y translation of that locator to a tracker node. Bring that tracker into shake; copy paste that data into a stabilize node and use that for stabilizing the camera.
it sounds complicated but it really isn't.