Stockertim Dictionary · Artificial intelligence
Zero-shot prompting
Also called zero-shot, zero-shot learning
In short
Zero-shot prompting means asking a model to carry out a task using instructions only, without including any worked examples of the inputs and outputs you want.
- Beginner
- Process
- Checked
In detail
In a zero-shot prompt the model gets a description of the task and nothing else to copy. It relies on patterns learned during training to understand what is wanted. For common, clearly described tasks such as summarising, translating or answering a factual question, this often works well and keeps the prompt short.
Zero-shot prompting struggles when the task uses an organisation's own categories, a particular house style or a format the model has not seen. Without examples, the model guesses at boundaries: whether a message counts as a complaint or a query, or how terse a label should be. Those guesses vary between runs and between models.
The usual approach is to start zero-shot, measure results on a small test set, and add examples only where the errors show the instructions are not enough. That keeps prompts as simple as possible while still meeting the quality needed, and it shows exactly which boundaries the examples have to teach rather than adding examples by habit.
Worked example
Northwind Bikes wants customer tickets labelled as repair, refund, delivery or other, and tests a zero-shot prompt on 20 tickets.
- Zero-shot prompt: "Label each ticket as repair, refund, delivery or other."
- Correct labels: 15 / 20 = 75%
- Most errors are tickets asking for a part to be replaced, which the team counts as repair but the model labels other.
The instructions alone reach 75%; the errors point to one boundary that examples or a clearer definition could fix.
In practice
Staff use zero-shot prompts for everyday drafting and summarising. For repeated classification or extraction, teams test a zero-shot version first and move to few-shot prompting if accuracy falls short.
Watch out
- Zero-shot does not mean zero context; the instructions can still be detailed.
- A good zero-shot result on a few cases says little about performance on hundreds.
Learn it properly
Official sources
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile · National Institute of Standards and Technology
Facts last checked on 3 November 2026.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.