Examples over Rules

The instinct, when a model gets something wrong, is to add a rule. It works — for that case. Do it fifteen times and you have a prompt that spends most of its budget on edge cases, degrading attention on the ninety percent of traffic that was never in question.

The prescribed replacement is a small, curated set of diverse canonical examples. For a model, a few well-chosen demonstrations act as pictures worth a thousand words: they carry formatting, reasoning pattern and implicit constraints at once, without any of it being spelled out.

The word doing the work is diverse. Five examples that differ only in their nouns teach one case at five times the price. Each example should earn its slot by demonstrating something the others do not.

A useful test: cover the corners, not the centre.

The typical case usually needs no example — the model already does it. What needs demonstrating is where your domain deviates from the default: the case that looks like it should be handled one way and is not, the input that should produce a refusal, the shape that is easy to get subtly wrong.

Three or four examples chosen that way outperform a dozen chosen by sampling. And each one should show the output as well as the input, because the output format is usually the thing you were trying to teach.