Why is few-shot prompting often more effective than a purely descriptive instruction?

A descriptive instruction hands the model a specification and asks it to imagine the output. An example hands it the output. That is the whole difference, and it matters most when the pattern you want is arbitrary — a particular date format, a summary that always leads with the verdict, a table whose third column stays blank when the value is unknown. None of those follow from anything the model knows about your domain; they are just how you want it done, and a paragraph describing them is a worse carrier than one worked pair.

Description also fails quietly. Where your wording is ambiguous the model picks a reading and produces something plausible, and you find out on the tenth output that it picked the other one. An example settles the reading before the first output.

The limit is worth knowing too. Examples steer form far better than they steer facts, and a set that is too narrow teaches an accident of your samples rather than the rule. Keep them close to the real input and vary them enough that there is nothing else to copy.