PRODUCTION SCENARIO
Your ticket-triage prompt carries eight examples, each already wrapped in example tags, and all eight are billing complaints labeled billing. The label set is billing, shipping and login, roughly even in production, with edge cases that sit between two of them. In production the model labels shipping delays and login failures as billing as well.
How should the examples be revised?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Replace them with three to five examples spanning the labels and the edge cases
Examples have to be diverse enough that Claude does not pick up an unintended pattern, and a set drawn from one label teaches exactly the wrong one; three to five relevant examples covering the label set and its edge cases is the documented shape. Adding one example per missing label is right when the set is already balanced and a label is simply absent, and it leaves an eight-to-one-to-one imbalance against a roughly even distribution. Written definitions are right when labels follow policy rules rather than resemblance, and they give up the most reliable steer. Tagging is right when examples are read as instructions, and these are tagged already.