Construct a useful bad/good pair
Use one task and the same known facts. Show a failure-producing instruction, then change the instruction that addresses that failure. Explain what to inspect and where the method stops applying.
Example
Task: classify a support message for a downstream consumer that accepts four categories.
Bad: Classify the message and explain your answer.
Good: Return JSON with category in billing|technical|feedback|other and a short reason. If the message lacks enough information, use other and identify the missing fact.
Expected: parse the JSON and validate category membership. This checks the output contract; it does not establish semantic classification accuracy. Evaluate accuracy with representative labelled requests separately.