When should I stop using it?

When output is broadly correct and consistently the wrong shape, or when results vary between attempts in ways you struggle to describe. Both are signs that a convention you hold is not in the request, and showing two examples fixes in a minute what more description will not.

Does a better model remove the need for examples?

It removes some of it. Newer models infer a great deal more from a well-worded request than earlier ones did, which is a real improvement. What no model can infer is the convention nobody stated, so the remaining gap is the same gap and the remedy is unchanged.

Is it worth naming something everybody already does?

Only because it makes the alternative visible. People who have never heard the term rarely think to add examples, and people who have tend to reach for them at exactly the right moment. That is the whole practical value of the vocabulary here.

Glossary

Zero-shot prompting

Zero-shot prompting means asking for something with no worked examples at all, relying entirely on what the model already learnt and on how clearly you described what you wanted.

In plain terms

It is what you have been doing all along. Type a request, get an answer. The term exists because there is a named alternative, and knowing there is one is the whole value of the name: it turns the thing you do by reflex into a choice you can notice making.

Why it matters

It works remarkably well and it is the right default, so the useful knowledge is not how to do it but when to stop. Its failures have a signature: answers that are broadly right and consistently the wrong shape, or that vary between one attempt and the next in ways you cannot describe. Recognising that signature is what tells you to reach for examples; missing it leaves the tempting conclusion that the tool is not very good.

How it works

You supply an instruction and the model produces something from what it already has. Everything unstated is filled in from what it learnt was typical, which is why the same request gives different people quite different results depending on how much of their context reached the request.

Modern models are far better at this than early ones, which changed the practical advice considerably. Tasks that once needed examples now work from a well-worded request, and the honest position is that this handles most ordinary work perfectly well.

Its weakness is anything with an unstated convention. Format, tone, level of detail and what to leave out are exactly the things people find hardest to describe and easiest to demonstrate, so those are where output comes back nearly right and slightly off in a way that is difficult to articulate.

It costs nothing beyond the request itself, which matters more at volume than anywhere else. A step running thousands of times a day sends no examples, occupies less of the budget and is quicker, so where it does the job well enough it is straightforwardly the better arrangement.

Which of the two failures are you looking at?

Both columns describe a disappointing answer and only one of them responds to examples. If the output is broadly correct but the wrong shape, and you find yourself struggling to say exactly what is off, that difficulty is the diagnosis: you are holding a convention you have never articulated, and two instances of it will convey what several attempts at describing it did not. If the output is confidently wrong about your prices, your policy or your customer, no amount of pattern-matching helps, because the model was never given the material and examples of the right format will not supply it. That case needs the information in the request instead. Telling the two apart takes seconds and saves the very common cycle of adding more and more examples to a problem that was never about the shape.

Seen in the wild

Ask for something ordinary with no examples and no preamble, which is the default arrangement and works better than the vocabulary around it suggests.
ChatGPT
Take a request that keeps coming back nearly right, add two real examples, and see whether the difference is the thing you have been failing to describe.
Claude
Run a high-volume step without examples where it performs well enough, since sending nothing is faster and cheaper on every single run.
n8n

Common misconceptions

People assume

It is the beginner's version.

In fact

It is the right default and handles most everyday work well. Examples are a specific remedy for a specific problem rather than a general upgrade, and adding them to a request already working costs space and speed for nothing. The skill is noticing which situation you are in.

People assume

If it fails, the model is not capable enough.

In fact

Usually the request left something unstated. A model with no examples fills gaps from what it learnt was typical, and typical is rarely your convention. Before concluding a tool is inadequate, try showing it two instances of what you meant, which takes a minute and settles the question.

Telling them apart