In a cognitive model (production system), they will be derived from psychological hypotheses about the contents of LTM: rules acquired from earlier learning. See next week's examples.
In an expert system, they will not correspond nearly as closely. The objective is to get a job done, not to build an accurate cognitive model. However, they will often be derived as descriptions of the behaviour of one or more human experts: by taking protocols, conducting interviews, setting classication tasks, etc. This is known as knowledge engineering or knowledge acquisition. Properly done, it involves a lot of psychological expertise, to make sure that you describe what the expert does rather than what he thinks he does.
Note that many statements that don't look like rules can be formulated as such. For example: ``all mammals have fur'' means ``If something is a mammal, then it has fur''. ``Lockjaw is fatal'' and ``A patient with lockjaw will die'' both mean ``If the disease is lockjaw, then it is fatal''. A small part of the knowledge engineer's job is recognising the hidden structure in such statements.