Zero-shot prompting

One instruction, nothing else

Zero-shot prompting means giving a model an instruction and nothing else — no example inputs, no example outputs, no demonstrated pattern to copy. The model has to infer, purely from what it learned during training, what a correct response looks like. “Zero” refers to the number of worked examples in the prompt, not to the amount of context: a zero-shot prompt can still include background information, constraints, or a persona. It just never shows the model what a finished, correct answer actually looks like.

A concrete zero-shot prompt:

Classify the sentiment of this review as positive, negative, or neutral. Review: “The battery life is fine, but the screen cracked after a week.” Sentiment:

No example of a correctly classified review appears anywhere. The model has to recognize “classify sentiment” as a familiar task shape and infer, from the instruction and the review text alone, which of the three labels applies and in what format to return it.

Why it works at all

Modern instruction-tuned models have seen enormous numbers of task descriptions paired with correct completions during training and instruction-tuning, so a request phrased in a conventional way often maps cleanly onto a pattern the model already knows how to execute. Sentiment classification, summarization, translation, simple extraction, and rewriting are common enough that zero-shot usually gets you most of the way there with a capable model. This is also why zero-shot is the right place to start: it costs the fewest tokens, takes the least prompt-engineering effort, and if accuracy is already good enough, every example you would add to a few-shot prompt is pure context-budget spent for no gain.

Where zero-shot breaks down

Three failure modes show up repeatedly.

Unusual output formats. If you want the sentiment label returned as a single lowercase word with no punctuation, say so explicitly. A prompt that just says “classify the sentiment” without specifying the output shape gets back whatever format is statistically typical for sentiment-classification answers in the training data — a full sentence, a JSON object, or a label with inconsistent casing from one call to the next.

Ambiguous task framing. “Summarize this” is ambiguous along several axes the model has to guess at: how long should the summary be, should it preserve technical jargon or simplify it, is the reader someone who already read the source or someone deciding whether to? Each guess is a chance to guess wrong, and with no example anchoring the prompt, nothing corrects a bad one.

Domain-specific or house-style conventions. A model has no way to know that your team labels tickets “P0/P1/P2” rather than “urgent/high/low,” or that your style guide bans the Oxford comma, unless you tell it — and even then, an instruction like “use our house style” without specifying what that means gives the model nothing concrete to follow.

Writing a zero-shot prompt that holds up

The fix for all three failure modes is the same: be as explicit as possible about the output you want, even though you are not demonstrating it. State the exact format (“respond with one word: positive, negative, or neutral”), resolve ambiguity up front (length, audience, tone), and spell out any house-specific conventions rather than assuming the model will infer them. A well-specified zero-shot prompt often outperforms a sloppy few-shot one, because explicit instructions remove guesswork that implicit examples sometimes leave only partly resolved.

The moment a task’s output shape is easier to show than to describe in words — a particular table layout, a specific tone, a format with several interacting rules — that is the signal to stop refining the instruction and add a worked example instead. That is exactly where few-shot prompting, next, picks up.