I usually go about it differently than hold off coding. I do try to understand the problem fully before I implement anything. I try to figure out what abstractions make the most sense and what to optimize for. Then I write a preliminary implementation and almost without fail I learn something new that changes the parameters of the "problem". Scrap it and do it again just with more knowledge. No amount of analysing and understanding up front have produced a better or faster result for me.