TokenPad

Prompting

Zero-shot prompting

Zero-shot prompting asks the model to perform a task with instructions only, supplying no worked examples.

Modern instruction-tuned models handle a great deal zero-shot, and it is the cheaper option — no examples means no example tokens on every request.

It becomes unreliable when the output format is unusual, when the task boundary is genuinely ambiguous, or when the model must adopt a specific style. Those are the cases where examples earn their tokens.

In practice

Zero-shot is the right default for tasks a model has plainly seen a million times — translate, summarise, classify sentiment. It stops being adequate the moment your task has a house-specific output format or a category boundary that is yours rather than the world's. That is the point where two examples cost 200 tokens and save an afternoon.

Common questions

When is zero-shot enough?

For tasks the model has plainly seen a million times — translate, summarise, classify sentiment — and where the output format is obvious. It stops being adequate the moment the format or the category boundaries are specific to you.

Is zero-shot cheaper than few-shot?

Yes, by whatever the examples cost on every request. Two 100-token examples at a million requests a month is 200 million input tokens — worth paying when they fix a real failure, wasteful when they were added defensively.

Measure it

More in Prompting