I think there are times when this works and there are times when it will be a huge waste of time. In general, I think it's often very hard to tell which approach will be more efficient for you. As some sibling comments have mentioned, I think the way each person's individual brain works is a significant factor, but I think there are other potentially unknowable factors that are significant as well.
The best thing I've come up with to deal with this is to use an iterative-deepening-like approach. There's a reason this algorithm (pre AlphaGo) was the most common approach for many game playing programs. The general idea is to go a ways down a particular path but always keep in mind some notion of the global suitability of this path and when it starts to look too hard, back up and investigate some other approaches to at least a shallow (but a little bit deeper than before) depth. This lets you avoid potentially costly dead-ends for relatively low overhead.
(These thoughts inspired from this nice talk: https://www.youtube.com/watch?v=Z8KcCU-p8QA)