What you are referring to is the "Topological Naming Problem", where the names of faces changed after operations (because OCC, the CAD kernel, doesn't name them at all!), causing mapped features to move or break.
There were always workarounds that made it possible to do what you say: you could offset your sketches from base planes rather than attach them to generated geometry, and otherwise avoid using generated geometry by making use of parametrics.
Once you knew how it wasn't particularly onerous. For a couple of years of 0.20 and 0.21 I had no TNP-related breakage because of this.
(There could be some problems with assemblies that were a bit more challenging.)
However, this is now essentially a solved problem in 1.0. There's a high level of mitigation in a stable face-naming algorithm that keeps track of the changes OCC creates. There are still edge cases being reported, but the implementation works well. I've stopped worrying about applying the normal base-planes technique.
(There are still edge cases of TNP in other CAD packages, even those built on Parasolid, because it's a research-grade problem; once you think about it in depth, it can never be "solved". It can only be mitigated to the point where the package usually does what you would expect.)
You can now sketch on faces, edit dependencies, make Pads and sweeps from solid geometry edges as well as sketch edges, refer to generated geometry in sketches, and see the subsequent features stay where they should be.
So this "main problem with FreeCAD" is essentially in the past, as it has been for a long time in the RealThunder branch where the mitigations were originally developed.
Another "main problem" -- the lack of a core Assembly workbench -- is also addressed.
The final "main problem" of FreeCAD, IMO, is that OCC's implementations of fillet, chamfer and thickness are less capable than other modern CAD systems; there are limitations regarding thick fillets on thin edges, essentially. It might be possible to get these improved upstream, or it might not be, but there are other ways to get fillets into your design (such as adding them in the sketches, where arguably any "functional" fillet should be anyway).
There were always workarounds that made it possible to do what you say: you could offset your sketches from base planes rather than attach them to generated geometry, and otherwise avoid using generated geometry by making use of parametrics.
Once you knew how it wasn't particularly onerous. For a couple of years of 0.20 and 0.21 I had no TNP-related breakage because of this.
(There could be some problems with assemblies that were a bit more challenging.)
However, this is now essentially a solved problem in 1.0. There's a high level of mitigation in a stable face-naming algorithm that keeps track of the changes OCC creates. There are still edge cases being reported, but the implementation works well. I've stopped worrying about applying the normal base-planes technique.
(There are still edge cases of TNP in other CAD packages, even those built on Parasolid, because it's a research-grade problem; once you think about it in depth, it can never be "solved". It can only be mitigated to the point where the package usually does what you would expect.)
You can now sketch on faces, edit dependencies, make Pads and sweeps from solid geometry edges as well as sketch edges, refer to generated geometry in sketches, and see the subsequent features stay where they should be.
So this "main problem with FreeCAD" is essentially in the past, as it has been for a long time in the RealThunder branch where the mitigations were originally developed.
Another "main problem" -- the lack of a core Assembly workbench -- is also addressed.
The final "main problem" of FreeCAD, IMO, is that OCC's implementations of fillet, chamfer and thickness are less capable than other modern CAD systems; there are limitations regarding thick fillets on thin edges, essentially. It might be possible to get these improved upstream, or it might not be, but there are other ways to get fillets into your design (such as adding them in the sketches, where arguably any "functional" fillet should be anyway).