Classify text into a custom set of labels two ways: once with a plain instruction (zero-shot) and once with a few labeled examples in the prompt (few-shot). Compare how the examples change the model's output format and accuracy.

Few-shot prompting is the fastest way to steer a model toward a specific output format or a task it does not handle well by default. Showing two or three worked examples often does more than a paragraph of instructions, because the model copies the pattern it sees rather than interpreting a description.

Requirements

Expected Output

Both approaches usually get easy cases right, but the few-shot version is more consistent about returning just the label.

Solution

The few-shot version carries its examples as alternating user and assistant turns, which is how you show the model a worked pattern in chat format. Those turns demonstrate both the mapping (this kind of message gets this label) and the exact output shape (a bare label, not a sentence). That demonstration is usually a stronger signal than the instruction alone, which is why few-shot output tends to be more consistent and easier to parse.

Asking a model to show its reasoning before answering often improves accuracy on multi-step problems. Pose a small arithmetic word problem two ways, once demanding only the answer and once asking it to think step by step, and compare the results.

Requirements

Expected Output

The direct answer is sometimes wrong; the step-by-step version reasons to the correct $50.

Solution

Reasoning takes tokens, and a direct answer gives the model no space to work through the steps, which is where multi-step problems go wrong. Adding "think step by step" lets the model lay out intermediate results before committing, and the explicit "Answer:" line gives you a stable place to parse the final value from. The cost is more output tokens and latency, so you reserve chain-of-thought for tasks where the reasoning actually changes the answer.