zero-shot prompting

basicwidespread

A request to a language model where you only describe the task without providing examples, and the model solves it using its existing knowledge.

In plain terms

When you type into a chatbot, 'Summarize this article in three sentences', you wrote a zero-shot prompt. You stated what you want, but did not show an example of what a good answer should look like.

It is like telling an experienced chef only: 'Make a light summer soup.' You do not give a recipe or show a photo because you trust that they have cooked enough to understand what you mean.

The term 'zero-shot' literally means 'zero shots', or zero example cases. The opposite is few-shot prompting, where you include a few worked-out examples in the request so the model sees the desired format and style.

For most everyday tasks, this is the simplest and fastest method, and it is generally sufficient for today's models. Examples are useful when the response format or tone is unusual or highly constrained.

How it works

You write the request: describe the task, and if necessary, the circumstances (who it is for, how long it should be, in what format you need it).
The model breaks the request into text fragments (tokens) and, based on patterns from the vast amount of text seen during training, identifies the task type.
Since it receives no example, it infers from the task description what kind of response fits and writes the probable continuation token by token.
If the result is not good, you can refine the instruction or provide a few examples – in this case, you are using a few-shot prompt.

An everyday example

You paste the long minutes of a condominium general meeting into a chatbot and ask: 'List in bullet points what was decided and how much it costs me.' You give no example, yet the model understands the task and returns a usable list.

Who is affected

everyoneAnyone who uses a chatbot is, in most cases, writing a zero-shot prompt without knowing it.

employeeFor writing letters, summaries, and notes, this is the fastest method, and if you know when it falls short, you save time. → Work & HR

parent, student, teacherFor learning, asking for explanations, and creating tasks, a clear instruction is sufficient, but the answer must be checked. → Education & Family

developerWhen building an AI agent or application, you must decide whether a simple instruction is sufficient or if examples are needed to stabilize the output. → Code

A common misconception

MythZero-shot means the model solves something it has never learned about before.

In reality"Zero" only indicates that the request contains no example. The model works from this knowledge, having seen countless similar texts and tasks during its training.

Limits and risks

If the request is vague, the model guesses: it can give a response of completely different length, style, or structure for the same sentence.
With strictly constrained formats (e.g., tables, data fields, corporate templates), it makes errors or deviates from the requested shape more often than without examples.
In a narrow field or under company-specific rules, the model cannot know what you expect unless you describe it or show it.
Even a confident-sounding response can be wrong, so always verify facts, numbers, and references.

Similar, but not the same

few-shot prompting – There, you include several worked examples with the request, whereas zero-shot provides none.
fine-tuning – Fine-tuning continues training the model itself on new data. Zero-shot prompting, however, depends only on how the request is phrased; the model remains unchanged.
zero-shot learning – In machine learning, this originally meant that the model recognizes categories from which it saw not a single example during training, for instance based on descriptions. Zero-shot prompting is looser than this: it simply means there is no example in the request. The model may have seen the task type many times during training.

Related concepts

Broader

Prompting, prompt engineering

Narrower

zero-shot chain of thought

On the edge

few-shot prompting, one-shot prompting, in-context learning, chain of thought, system prompt, large language model

A brief history

The term 'zero-shot' originates from machine learning research. Since the late 2000s, it has denoted a model handling a task or category it was not specifically trained on. For language models, OpenAI's 2020 GPT-3 study widely popularized the distinction between zero-shot, one-shot, and few-shot.

Where things stand

At the end of 2026, this is the default way of using language models: models handle most everyday tasks well without examples. Examples are mainly added for fixed-format tasks or tasks requiring a very distinctive style.

Status: 2026 November

Also known as

zero-shot prompting · zero-shot prompt · zero shot prompting · 0-shot prompting · zero-shot prompting · zero-shot prompt · zero-shot prompting · zero-shot instruction

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.