A system prompt describes the task in words. But some things are easier to show than to describe: how short a reason should be, when a tender is “partly suitable”, which score goes with which judgment. In this lesson, the model gets examples for that: finished assessments, the way you’d like them. At the end, you use the local model to measure whether the examples really make the Kladde’s assessment better.
Without and with examples
So far, the model gets only the instruction and the tender. That’s called zero-shot: It has to solve the task without a single example. With one or a few examples, it’s called one-shot or few-shot.
The model doesn’t permanently learn anything from them; it isn’t changed. The examples work only in that one request: They make it more likely that the answer looks like the ones shown. In the next request, they have to come along again.
Two things you should know about this:
- With the chat interface, Microsoft puts examples behind the system message as pairs of a user message and an assistant message. The Kladde does the same.
- Microsoft’s page on prompt techniques says right at the start that these techniques aren’t recommended for reasoning models such as gpt-5 and the o series. So whether examples help depends on the model. That’s why you measure it in exercise 3 instead of assuming it.