You’ll likely be interested in the Dynverse and its associated publication, which reviews a whole bunch of trajectory inference methods to identify which work best for given situations.
As for choosing the starting/end point, it’s hard to get around doing so manually. RNA velocity will try to pick a good starting point, but for trajectory inference/pseudotime, you’ll probably have to define it manually. Which is fine, trajectory inference is basically just a fancy way of ordering cells along a line. If you have biological expertise such that you know which cluster is your progenitor population, why wouldn’t you use it?
As for whether it’s better to work directly from your Seurat object or run the whole thing through Monocle, this is one reason I prefer slingshot, as it just needs embeddings – it doesn’t care where you get them from. Coupled with tradeSeq, it makes looking at expression changes along a trajectory fairly easy. Really though, it shouldn’t matter if you use Monocle or Seurat for the pre-processing, they should yield relatively similar results