Speech or writing can have various levels of redundancy. Less redundancy is more efficient (and thus better) by relying on the audience's inferences. More redundancy is more robust to misunderstanding. There is some optimal amount, giving the most efficiency possible with an acceptably low rate of misunderstanding.
I find that many people speak much more redundantly than optimal. To reduce this, I propose a criterion to suggest the optimum: speak with sufficiently little redundancy that the listener cannot (semantically) predict what you say, but no further. If you get to a point where you could increase redundancy but remain semantically unpredictable, you probably went too far.
In conversation with me, it is especially easy to test this criterion, and especially desirable to obey it. I finish other people's sentences a lot, being intelligent and overconfident in my prediction. Thus, if you make me unable to predict the rest of what you're saying, I would give up on trying to do so, which you can notice and would probably prefer.
The criterion may be harder to test with others, but you might ask for predictions on the gist of what you'd say next; if they're accurate, your next statements are redundant, perhaps indicative of a general excess redundancy.
Adjust how much to omit based on the concentration and domain-intelligence of the listener. Your starting point should probably err more on the side of "omit more redundancy" than it currently does.
This also applies to writing. In either speech or writing, tho, when your audience is multiple people, adjust expected audience-inference to somewhere between that for the median audience-member and that for the least-inferring audience-member. Hence this does not apply so much to lectures or writing for large audiences. And that's why this essay is so much more verbose than I seem to be advocating.
Communicative efficiency and robustness are a trade-off with an optimum.
Optimally, the listener can't predict the meaning of what you say.
I actively predict what you say, so it's easy to optimise that way with me.
Explicitly ask, with others.
Large audiences lower inference standards.