學術辯論與觀點比較技能

純指令技能。LLM 結構化思考 + read_draft, search_local_references, mcp_cgu_deep_think, mcp_cgu_spark_collision。

觸發：辯論、pro/con、devil's advocate、挑戰假設、compare viewpoints

---

Framework 1: Academic Debate（雙方辯論）

針對研究議題結構化正反方論點。按證據等級排列： SR/Meta > RCT > Cohort > Case-Control > Cross-Sectional > Case Series > Expert Opinion

輸出：Position A (supporting) → Position B (opposing) → Methodological Considerations (偏誤) → Synthesis (agreement / disagreement / clinical bottom line)

---

Framework 2: Devil's Advocate（魔鬼代言人）

系統性挑戰一個研究主張。5 面向：

1. Methodological — 研究類型對應偏誤（見下方）
2. Statistical — multiple comparisons, effect size, CI width, missing data, power
3. Generalizability — population, setting, timeframe, intervention fidelity
4. Alternative Explanations — confounding, reverse causation, temporal trends, Hawthorne
5. Likely Reviewer Questions — 基於 study type

產出：Strengthening Recommendations（address counter-arguments, add sensitivity analysis, cite SR, frame conservatively）

---

Framework 3: Viewpoint Comparison（觀點比較）

多個理論/方法系統性比較。預設比較維度： Evidence Base, Theoretical Foundation, Clinical Applicability, Patient Safety, Cost-Effectiveness, Current Guidelines, Limitations

---

Study-Type-Specific Biases

Type	Key Biases
RCT	Selection, performance, detection, attrition, reporting
Cohort	Selection, confounding, information, loss to follow-up, healthy worker
Case-Control	Recall, selection (controls), confounding, misclassification, temporal ambiguity
Cross-Sectional	No temporality, prevalence-incidence, non-response, information
Retrospective	Information (records), survivorship, confounding, missing data

Auto-detect：randomized/RCT=RCT, cohort/prospective=Cohort, case-control/odds=Case-Control, cross-sectional/prevalence=Cross-Sectional, retrospective/chart=Retrospective

---

使用原則

1. 基於證據，引用已存文獻
2. 平衡呈現雙方
3. 偏誤分析對應正確研究類型
4. 最終給臨床建議（為 Discussion 服務）