Suppose you ask a language model for short customer utterances about login problems, but it keeps writing full sentences. Adding more instructions may not help. A single example—“Forgot password”—shows the shape of the answer you want. section 7.2.2
One-shot prompting puts one demonstration in the prompt to guide the model’s next response. The “shot” is the example: zero-shot provides none, one-shot provides one, and few-shot provides several. An example can convey a format or grammatical pattern more clearly than instructions alone. section 7.2.2
Generate a different kind of utterance
To generate login-related sentence fragments, request fragments and give “Forgot password” as the first item. The model can then continue with phrases such as “Unable to login” and “Account locked out.” The example demonstrates what counts as a fragment; it does not guarantee that every generated item will fit the intent. section 7.2.2
Guide a handoff summary
A chatbot handoff needs a brief account of why the user requested a representative. Instructions alone produced a speculative claim-denial summary even though the conversation said the claim was paid. A one-shot prompt pairs a sample transcript with a suitable summary, then supplies the transcript to summarize. Check the result against the conversation: the example improved its structure but did not eliminate speculation. section 12.3.1
Choose an example that matches the output you need, then inspect the result for accuracy and usefulness. Prompt formatting can matter: the summarization example needed distinct markers to separate instructions, transcripts, and summaries. For generated intent data, refine unsuitable outputs rather than assuming the first response is ready to use. section 12.3.1 section 7.4
