Create Near-Miss Examples That Teach Your Twin When Advice Does Not Apply

ROLE You are a contrastive learning dataset designer. OBJECTIVE Create examples that teach my AI twin where a recommendation applies and where a superficially similar situation requires a different response. CONTEXT My recommendation or method is [RECOMMENDATION]. Its known applicability conditions are [APPLICABILITY CONDITIONS]. Real examples I can share are [SOURCE EXAMPLES]. INSTRUCTIONS Identify the smallest contextual differences that would change the recommendation. Construct paired scenarios that remain similar except for one decisive variable. For each pair, provide the appropriate response, the evidence that changes the response, and the tempting but incorrect generalization. Include cases where the right behavior is to ask a question rather than advise. Clearly distinguish source-derived examples from synthetic teaching examples. Propose a consistent record structure and explain how to keep a separate set of analogous cases for evaluation. CONSTRAINTS Do not invent exceptions and present them as my established beliefs. Mark proposed exceptions for my approval. Avoid fabricated statistics and flag assumptions. Do not place sensitive real-world details in synthetic cases. OUTPUT FORMAT Applicability summary, six contrastive pairs, record schema, and expert-review questions.