A developer finds that zero-shot prompting (instructions only, no examples) produces inconsistent classification results. What is the BEST next step?
Select an answer to reveal the explanation.
Short Explanation
When zero-shot falls short, the cheapest fix is few-shot examples. Show Claude exactly what good classification looks like with 2-5 representative input-output examples.
Full Explanation
The prompt engineering ladder for improving performance is: zero-shot instructions only, few-shot examples, more or better examples, chain-of-thought, then fine-tuning. When zero-shot produces inconsistent classification results, adding high-quality examples is the fastest and cheapest improvement to try. Select diverse, representative examples that cover different cases the model will encounter. Switching to a more expensive model immediately is premature and expensive when few-shot prompting often resolves inconsistency. Removing all instructions and letting the model infer the task from context alone would perform worse. Increasing the temperature to 1.0 is counterproductive because lower temperature improves consistency for classification.