Divide the various things you want to say into three or four messages. (Three is, perhaps, slightly better than four, and either is much better than any other number.). A message might consist, for example, of some important point together with supporting arguments and examples, or of a collection of remarks which share some common property. Each of these messages will, eventually, become a short talk in its own right. Assign to each message a title (e.g., ``The Initial-Value Formulation of General Relativity''), and invent, for each, a non-technical summary of the message in a sentence or so (e.g., ``The `initial time' becomes, in general relativity, a spacelike three-dimensional surface; the `state of the the gravitational field at that time' becomes a pair of tensor fields on this surface, subject to certain constraint equations; the `evolution of the system' is then described by equations which give the change in these fields under changes in the spacelike surface.''). The more cohesive each message is, the better.
It is almost always necessary, in order to obtain such an organization, to recast the subject into a form which is essentially different from the way in which you normally think about it. (In one's mind, the subject is innumerable interconnected small points. In the talk, the subject will be three of four main points.). In particular, one often has to omit some details one might otherwise have wanted to say, omit connections between certain points, or add material to fill out a message. The idea is that, with three or four messages, the audience can grasp and hold onto the structure of the entire talk. (People simply are not going to come away from any talk with more than three of four essential points.). The difficult thing about planning a talk, in my opinion, is to divide things into messages which are sufficiently specific and cohesive that each can be treated as a unit (hence, remembered by the audience), and yet sufficiently general that, taken together, the messages tell one's story. One would like to come up with several (hopefully, very different) organizations, and then select the best for further refinement.