모델별 답변이 갈릴 때 뭘 먼저 확인할지 정하는 프롬프트

I asked the same question of several models and got different answers. **Work out what the difference means.**

1. **Separate where they agree from where they differ.** ***Agreement is not correctness*** — models trained on overlapping data repeat the same errors together. **Note that explicitly**
2. **Classify each disagreement:**
   - **Factual** — they cannot both be true. ***This one has to be resolved before use***
   - **Interpretive** — they read the question differently. **Then my question was ambiguous, not their answers wrong**
   - **Scope** — one covered more ground
   - **Emphasis** — same content, different weighting. Usually not a real difference
3. **For each factual disagreement:**
   - **Which side is checkable, and how**
   - ***Which answer is more specific*** — **and treat that as a warning, not as evidence.** The more detailed answer is often the more confidently wrong one
4. **Where only one model produced a claim the others did not mention at all** — flag it. ***A claim nobody else makes is either an insight or a fabrication, and the shape is the same***

**Output:** agreed / differs (with type) / **what to check first**.

⚠️ ***Do not decide which model is right.*** **Do not average them into a merged answer** — a blend of a right answer and a wrong one is wrong.