How many examples should I use?

Two or three, chosen for how well they represent what you want rather than for variety. Add a fourth only if it demonstrates something the first three do not. The number matters far less than whether they agree with each other on format, tone and depth.

Should the examples be real or made up?

Real, wherever possible. Genuine past work carries conventions nobody in your organisation has ever written down, and those are precisely the parts that make output feel right or wrong. Invented examples get you output that matches the invention, which is a quiet way of missing the point.

Is this cheaper than fine-tuning?

Enormously, and it is the thing to try first. It needs no dataset, no training and no waiting, and it can be changed in a minute. Fine-tuning earns its place when the same pattern is needed at a volume where paying for the examples on every single request starts to add up.

Glossary

Few-shot prompting

Few-shot prompting means putting two or three worked examples inside your request, so the model can match a pattern you have shown it rather than interpret a description of one.

In plain terms

Show, do not tell. Instead of describing the format, the tone and the level of detail you want, paste two examples of it and add the new case. The name sounds technical and the practice is what anybody would do when briefing a colleague: here are two we did earlier, now do this one.

Why it matters

It is the highest-return habit available in this whole subject, it costs nothing, and a great many people never form it. Describing a house style accurately is genuinely hard and takes paragraphs; showing two instances of it takes seconds and conveys more. Where output keeps coming back nearly right in a way you struggle to articulate, this is almost always the fix, and it is a great deal cheaper than the training people reach for instead.

How it works

The examples go inside the same request as the task, above the new case. The model treats them as part of what it is continuing from, so the shape they establish carries into what it produces next. Nothing is stored and nothing is learnt; the effect lasts exactly as long as the examples are in front of it.

Consistency in the examples matters more than the number of them. Two that agree on format, tone and level of detail work better than six that vary, because inconsistency teaches that variation is acceptable. Picking the two that best represent what you want beats gathering more.

Where the examples are drawn from decides what you get. Real cases from your own work carry conventions you would never think to describe, including the ones nobody has written down. Invented examples produce output matching your invention rather than your practice, which is a subtle way of getting exactly what you asked for and not what you wanted.

It costs space on every request, which is the trade. Examples are sent again each time and compete with your documents and your conversation for the same fixed room. That is nearly always worth it for the quality gained, and it is a reason to use two good examples rather than eight adequate ones.

The same request, with the examples above it

Four short parts, and the whole technique is visible in them. The task is stated once and briefly, because the examples are doing the work that a long description of tone and format would otherwise have to do badly. The two examples are real replies rather than invented ones, which matters more than it sounds: genuine past work carries the greeting, the typical length and the way your organisation signs off, none of which anybody has written down and all of which would be missing from something composed for the purpose. They agree with each other, so nothing in them suggests that variation is welcome. And the new case sits at the bottom, immediately after the pattern, which is where the model will continue from.

Seen in the wild

Paste two of your own past outputs above a new case and compare what comes back against your usual attempt at describing the style in words.
ChatGPT
Attach a handful of real examples alongside the task in an assistant that handles long material comfortably, so the pattern comes from your practice rather than your description.
Claude
Put the examples into an automation step so every run matches the same pattern rather than depending on who wrote the request that day.
n8n

Common misconceptions

People assume

The model learns from the examples.

In fact

It matches them for the duration of that request and retains nothing. Send the same task tomorrow without them and you will get your old results back. If the pattern needs to apply permanently, it belongs in standing instructions rather than in one conversation.

People assume

More examples give better results.

In fact

Consistency does. Two examples agreeing on everything that matters outperform six that disagree on tone or format, because disagreement teaches the model that both are acceptable. Adding examples is worth doing until they start varying, and worth stopping at that point.

Telling them apart