Same model. Just ask differently.

No training, no new weights — only the context changes. Classify the review: “took forever and the fries were cold.”

The prompt we send
What the model returns
Confidence in a usable answer

In-context learning: showing worked examples (few-shot) or stating the rules up front (a system prompt) steers the answer without changing a single weight. The amber block is the part of the context doing the work right now.