Zero-Shot vs. Few-Shot Prompting
Zero-shot prompting asks the model to do a task with instructions alone. Few-shot prompting adds a handful of worked examples before the real question, showing the model exactly what a good answer looks like — not through training, but through the context-window mechanism from lesson 1.
Predict before you look
Both prompts ask for the same JSON classification. Before you look: do you expect the two outputs to differ in which category they pick, in their format, both, or neither?
Both got the category right. The real difference is format: zero-shot wrapped its answer in a ```json code fence — reasonable for a chat interface, but it would break a naive JSON.parse() call in real code. Few-shot's two examples never used a code fence, so the model's output didn't either — clean, parseable JSON on the first try.
That's the honest, more common story with few-shot prompting: on easy tasks the content of the answer often doesn't change, because a capable model already gets it right. What reliably changes is format adherence — exact labels, exact structure, exact tone — because examples pin down conventions that instructions alone tend to leave the model guessing about.
In the transcript above, what did few-shot prompting actually fix?