How few-shot prompting works
Instead of describing the output you want, you show it. The request contains a small set of worked pairs — this input, that output; this input, that output — and then the real input with the output left blank. The model continues the pattern it can see, which is the thing it is best at doing.
Asking with no examples at all is often called zero-shot, and for straightforward jobs it is perfectly adequate. Examples earn their place when the requirement is a shape rather than a subject: a particular way of writing product descriptions, a fixed set of categories to sort enquiries into, a house style for meta descriptions, or the specific way your team phrases an apology.
Why few-shot prompting matters
It solves the problem that adjectives cannot. Telling a model to write in a tone that is professional but warm produces whatever that phrase averages out to across everything ever written. Showing it three pieces of your own writing communicates the tone precisely, including the habits you would never think to describe, such as sentence length or how you open.
It also makes output consistent enough to use in a process. If several people run the same task, or the same task runs every day, examples pin the format so the results can be compared, pasted into a spreadsheet, or fed into the next step without someone reformatting them each time. That consistency is usually worth more than any single output being brilliant.
Common mistakes with few-shot prompting
Choosing examples that are all alike is the first. If every one is a short product with a simple name, the model learns that shape and applies it to the complicated cases too. Include an awkward one — the long name, the item with a caveat, the enquiry that could fall into two categories — and show how you want it handled.
Using flawed examples is the second, and it is worse than having none. The model copies what it sees, including a typo, a missing detail or a claim you cannot support, and it will reproduce that fault across everything it touches.
The third is forgetting that examples are not free. They are sent with every request, so they consume budget and part of the context window each time. Long examples on a high-volume task get expensive quietly.
How to act on it
Pull the examples from work you have already approved rather than writing them for the occasion. Pick a handful that differ from one another, include at least one difficult case, and check each is genuinely correct before it becomes the pattern for everything that follows.
Keep them in a template with a blank for the real input, so the same set gets reused rather than reassembled. Trim to the fewest examples that hold the format, since each one costs tokens on every run. If you find yourself needing many examples for the same task repeatedly, that is the point at which fine-tuning becomes worth investigating — but only then, because a good template is far cheaper to maintain.
