Zero-shot Prompting

Two support leads, in an exercise constructed for this chapter, label the same 40 tickets using the ticket router's first prompt from Chapter 1. Working without conferring, they agree on 28. The prompt names six categories without defining any of them, so nothing in it says whether a duplicate charge with a lockout goes to billing or to account, whether a stated deadline makes a ticket high or urgent, or whether INV-20931 counts as an order number. Each lead fills those gaps with a sensible guess, and the guesses differ. A model reading the same prompt does what the leads did: where the instruction is silent or vague, it does not stop and ask. It falls back on what is typical, which may not be what your business means.