I think it depends on your geometric domain. What I mean is that if you're working with a flat plate, then the sketches are the way to go. If you're working with a 3D form, then there are likely some better options for you.
When I'm working with complex organic-ish 3D shapes, I'll often use the Form environment to create TSplines that are a close approximation to the generated mesh and then boolean them together with primitives (e.g. cylinders) that work as my interface geometry (e.g. bolt bosses). Another technique is to use a series of offset planes and the mesh-enabler to craft spline-based cross-sections that can then be lofted. In both cases, you're very hands-on with the geometry creation and as @johnsonshiue said, this is pretty iterative since the final form needs to be re-validated with FEA.
Autodesk Generative Design automates much of this process with its auto-SAT creation. While it's still somewhat early in its development, there is a ton of effort going into making that process even better. Where Generative Design starts to outpace Shape Generator is when you start exploring the entire design space and not a singular seed. When you have to create tens (or more) of these new forms at a time, the auto-creation is a life-saver. The SAT creation allows us to validate generated shapes quickly by tossing it right back into Nastran to confirm what Generative Design suggests is true. In the end, I typically still model our end solution in a parametric way - regardless of whether the form was synthesized by Generative Design or optimized by Shape Generator.
K. Cornett
Generative Design Consultant / Trainer