Debugging a Few-Shot Prompt That Stopped Improving

You're building a prompt that classifies support tickets into one of six categories. Zero-shot accuracy on your 200-example golden set is 74%. You add 3 few-shot examples (one per common category) and accuracy jumps to 89%. A teammate suggests adding 15 more examples, one for each remaining edge case they can think of, reasoning "more examples can only help." You add them and accuracy comes back at 86% — slightly worse than the 3-example version — while cost per call has roughly quadrupled.

1. Explain, mechanistically, why adding more examples made things worse here, considering at least two plausible causes.
2. Describe how you would methodically find the right example set, rather than guessing counts.
3. Your teammate now proposes fine-tuning on all 200 golden examples instead, to "bake in" the classification behavior permanently. Is that the right next move? Justify your answer.