Typically, yes, it's a matter of finding a known NP-complete problem that's somehow similar to the one you're trying to work with. So if you're dealing with a problem about formulas, you probably want to reduce some version of SAT or 3SAT to it.
For graph problems, you probably want to reduce some other graph problem. Problems about long paths and cycles probably come from Hamiltonian path/cycle. Problems about classifying vertices into types sound like colouring problems. Problems about graphs containing or not containing some structure might come from clique or independent set. Problems about dividing graphs in two might be Max Cut, or Subset Sum.
Reductions from one type of problem to another are typically more difficult. If you ahve to do that, think about how you can use your target problem to encode things that are needed in the known NP-complete problem. For example, when you reduce 3SAT to independent set, being in or out of the independent set corresponds to being true or false; when you reduce 3SAT to 3-colourability, the three colours you use are "true", "false" and "er, the other colour". But, in these cases, the reduction gadgets tend to be quite fiddly.
Another thing to bear in mind is that, if problem $A$ looks like problem $B$, which you already know to be NP-complete, it might be possible to modify that proof to make it work for $A$. For example, consider 4-colourability. The easy reduction is from 3-colourability: given a graph $G$, add a new vertex, connect that to everything and the new graph is 4-colourable if, and only if, the original graph was 3-colourable. But, if you didn't see that and you knew the reduction from 3SAT to 3-colourability well, it probably wouldn't be very hard to modify that reduction so the four colours were "true", "false", "er, the other colour" and "gee, there are a lot of colours today".