This is the prototypical problem in AI/ML known as feature-selection. If you do not have too many features, typically one would just use them all(with the exception of features that are known to to be correlated), in this case you would want to perform feature engineering wherever possible as well. One the other hand, if you have lots of features, and some are likely to be useless, you would use a feature-selection technique.
There are many many methods for doing this. One can simply use their intuitive grasp as their problem domain at the simplest level. However, there are also many algorithmic ways of performing it. Some examples of this are a low variance filter, or using a ensemble(classifier or regression depending on the problem) which can then be used to order features according to their derived importance(I personally like this).
A search of "feature selection methods in ML" will yield many potential ways to accomplish your goal.
As far as what to grab if you don't have many, as many as possible wrt domain knowledge.