Zero-shot prompting relies on the model’s existing learned capabilities and the clarity of the instruction. It is convenient for straightforward tasks because the prompt stays short and does not require example preparation. Results can be inconsistent when categories are subtle or the expected output format is unclear.
A request such as "Summarize this email in two sentences and name the sender's main request" is zero-shot: no sample summaries are included. It works well for common tasks that models already handle. If outputs vary in format or the labels are specific to your business, add a few examples or ask for structured output.
Start with a direct instruction and explicit constraints, then compare the result with a few-shot version if errors persist. Use a fixed evaluation set to determine whether examples actually improve performance. Related: Few-Shot Prompting, Prompt Engineering.