Zero-shot means asking for the task with no worked examples.
You rely entirely on what the model already learned in training.
It works well for common, clearly described tasks.
It is the fastest and cheapest thing to try first.
It becomes unreliable when the task has an unusual format or edge cases.
Zero-shot means asking for the task without showing a single example.
You describe what you want and rely on what the model already learned.
For common tasks described clearly this works remarkably well.
Translation, summarisation, classification into obvious categories and rewriting all suit it.
It is the cheapest option, since no example tokens are sent or billed.
It is also the fastest to write, so it is the right first attempt.
It weakens when the task has an unusual output format.
The model will produce something reasonable that is not quite the shape you needed.
It weakens again when your categories are specific to your business.
Urgent, blocked and waiting mean particular things in your team that the model cannot guess.
Edge cases are where it fails most quietly, since the common cases keep working.
The upgrade path is to add two or three examples and become few-shot.
Adding examples of the edge cases specifically is usually worth more than adding more examples.
Larger and instruction-tuned models are markedly better at zero-shot than base models.
Always try zero-shot first, then add examples only where it measurably fails.
One question on this card. Every option explains itself once you pick, right or wrong.
17 concepts in this module, then a graded test with every answer explained.