Zero-shot prompting
Asking a model to do the task without showing it any worked examples first. The prompt is instructions and input only.
See it
What it is
Zero-shot prompting means asking a model to do a task without putting any worked examples in the request. You give it the instruction, the input, and the output rules, then rely on what the base model already learned. It is the clean baseline against which one-shot and few-shot prompts earn their extra tokens. The ladder got its names from the 2020 GPT-3 paper 'Language Models are Few-Shot Learners', which reported zero-shot, one-shot, and few-shot scores side by side for every benchmark it ran.
Start here for familiar tasks such as summarising, rewriting, broad classification, or extracting obvious fields. A zero-shot prompt is cheap to write, cheap to send, and easy to change. If the model keeps missing a house style, confusing labels, or mishandling an edge case, that is the signal to add a few carefully chosen examples.
Gotcha: zero-shot does not mean context-free or instruction-free. The system prompt, retrieved documents, field definitions, and output schema can all still be present. It only means there are no demonstrations of the task. A vague sentence with no examples is still vague, not magically neutral.
Ask AI for it
Build a zero-shot text classifier on the OpenAI Responses API. Define the allowed labels as a JSON Schema enum and send it through Structured Outputs with `strict: true`, so the reply cannot drift off schema. Describe the boundary of each label in one sentence, and require exactly one label plus a confidence score from 0 to 1. Do not include any sample inputs, sample outputs, or worked demonstrations in the prompt. Add an 'unknown' label for inputs that do not contain enough evidence, validate the parsed response against the same schema in code, and return only the parsed object.