The Third Umpire · Lesson 4 — The Examples ← Course

The examples — few-shot prompting.

Describing the job is good. Handing over a couple of solved cases is better.

Last lesson you spelled the job out in words. There is a more reliable way to land the answer you want: show the model a few examples of the job already done, then ask. Asking cold, with no examples, is zero-shot; laying a handful of worked cases down first is few-shot.

Picture handing the umpire two earlier calls already settled — one given out, one not — each written up the same way. “Here is how we call these.” Faced with a fresh appeal, the umpire answers in the same shape without being told to.

It works because the model is, underneath, a pattern-continuer — the idea from Lesson 1. A worked example nails down the format, the level of detail, even the tone far more exactly than a sentence of description ever could.

None of it sticks. The examples do their job only while they sit in the prompt — start a fresh conversation without them and the model is back where it began.

So reach for this when the model nearly does what you want but lands it in the wrong shape. Two or three good examples teach by demonstration — awkward edge cases and all — where a paragraph of instructions only gestures.

Show the umpire zero, then one, then two prior calls — and watch the new verdict snap into the same shape.

No examples yet — zero-shot. The umpire answers in its own words.
prior calls shown first · examples
↓ then the real one ↓
the new appeal · to be ruled
the umpire's call

Examples first, then the question.

A few-shot prompt is literally that: the solved examples, then the real appeal, handed over together. The model continues the pattern it was shown.


      

Show, don't just tell.

Describing the format is good; demonstrating it is better — the model just continues the pattern you lay down. And the examples do more than fix a shape: add a contrasting one and the range widens. Show only an OUT and it copies OUT; show a NOT OUT alongside, at lower confidence, and now both verdicts sit in its repertoire to draw from.

Few-shot is the fix when the answer is close but mis-shaped. Show two or three clear, varied examples — the tricky calls included — and let the model copy.

Next: how many? →
Go deeper — getting few-shot right optional

What counts as a “shot”

One “shot” is one worked example — a single solved case. Show none and it is zero-shot; one is one-shot; a small handful is few-shot. The word just counts how many finished examples you place before your real question.

Where the examples sit

The examples ride in the prompt itself, alongside the system briefing and your appeal — the same context window from Lesson 2. They are part of the input for this one request, not stored anywhere the model keeps.

Order matters

The model leans hardest on the most recent examples, so put your clearest, most representative call last. If two examples disagree, the later one tends to win.

Examples now, or training later

Few-shot teaches a pattern for this one conversation, right in the prompt. If you need the model to always know the pattern without being shown each time, that is a different tool the course reaches near the end.