Rule #1: The sample should be sufficiently numerous and various.
A single instance is usually not enough to support a
generalization. In the absence of other knowledge, the fact that
a single S is P won't tell us whether S and P are connected, so
generalizing that all S are P would be hasty. We need to look at
a number of Ss; if all of them are P, then we have better
evidence of a connection.
Also, it's equally important to test a variety of Ss. How much
variety is enough? The general rule is that a sample of Ss
should vary in every property (other than being S) that might be
responsible for their being P.
Suppose you were considering buying a Toyota. You might ask
people who owned Toyotas whether they were satisfied. Suppose
all of them have had problems with the car. Does that indicate
that Toyotas are unreliable? This conclusion would be stronger
if you checked cars of different years, with different options,
bought from different dealers.