A helpful feature tells the model something meaningful about the example. It might describe recent usage, counts, scores, categories, time-based patterns, or other properties that are available at prediction time.
The key phrase is "describes the example." A good feature says something about the case the model is trying to reason about. It gives the model context instead of just giving the dataset one more column to carry around.
Good beginner feature judgment starts by asking whether the column genuinely describes the example in a way that could support prediction.
This is a practical question, not an abstract one. The learner should be able to say why a field helps rather than just trusting the fact that it exists.