I can't define any hard and fast rules there. Sometimes I draw the same model three times until I get a handle on how to make it happen. Based on measurements, yes, I often type in X, Y, and Z values. I also measure angles on actual parts, then rotate, or rotate/shear splines.
To keep calculations easier, I sometimes shift or rotate the whole model around a particular point I'm working on so that part of the model is at the origin on one or more of the axes.
I often draw parts of a single model as several separate models, then join them, adding detail or redrawing as I learn more about how their splines are meeting.
Eyeballing for some parts of models is sometimes sufficient, with tweaking based on measurements.
For bevels, I use the pythagorean theorem to calculate height and width (based on a measurement of the "hypotenuse"), and do a lot of scaling after calculating percentages.
I also tweak a lot of bias handles. Thank goodness for Emilio Le Roux's SetBias plugin and for Yves Poissant's tutorial on beveling (although I find that the magnitude values in Yves' tutorial for beveling a cube don't work for me any more. The values were great in earlier versions of Animation:Master, but I think they stopped working around v9 or v10. For example, I use a magnitude of 355 for a 4 cp circle. I have a whole spreadsheet full of bias handle values for different numbers of lathe cross sections. No, I didn't calculate them--I don't know how. I experimented with tweaking bias handles on peaked cps using as patterns smooth cps. I didn't see a clear pattern emerge that would have allowed me to calculate these values. The wizards at Hash could probably say something about that however.)