The question itself is an ironic one: the most "clever" way is rather to not use SKAN at all...
Somewhat useful as a secondary proxy to check attribution through another lense, but that's pretty much it
I use it as an additional data point to compare different scenarios but always having in mind that it underreports compared to reality (tested and proved it multiple times).
Sometimes it is useful because it reports more than probabilistic methodologies like AEM or ADC but to be really honest, it is not a decision maker for me nowadays