When an LLM produces the wrong kind of training utterance, adding examples may help—but a zero-shot prompt deliberately starts with instructions only. It specifies the task without showing an example of the desired input-output behavior. section 7.2.2 section 7.2.4
Generate login utterances without examples
A zero-shot request might ask an LLM to generate ten varied customer utterances for people who cannot log in, using nouns and verbs related to passwords and accounts. The output can include “I can’t log in,” “My password isn’t working,” and “My account is locked.” To demand fragments instead, add the instruction “Use sentence fragments only”; this remains zero-shot because no example fragment is supplied.
Use zero-shot prompting as a quick first pass, then evaluate its usefulness:
- Generate examples for an intent such as store location, varying nouns, verbs, and grammatical structures. section 7.2.4
- Compare the time and variety of zero-shot output with manually written examples and one-shot or few-shot prompts. section 7.2.4
- Add useful utterances to training or test data only after checking whether they improve classification accuracy and do not worsen balance. section 7.2.4
