devils-advocate

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Devil's Advocate Protocol

魔鬼代言人协议

Pre-commitment adversarial reasoning to prevent early lock-in and expose blind spots.

在做出承诺前进行对抗性推理，以防止过早锁定并暴露盲点。

When to Apply

适用场景

Activate this protocol when:

Choosing between architectural approaches
Selecting libraries, frameworks, or tools
Planning implementation strategy
Recommending one approach over alternatives
User asks "should I...", "what's the best way to...", "which approach..."
During
```
architect
```
,
```
Plan
```
, or
```
blueprint
```
workflows
Making trade-off decisions with non-obvious answers

在以下情况激活本协议：

在不同架构方案间做选择时
选择库、框架或工具时
规划实施策略时
推荐某一方案而非其他替代方案时
用户询问“我应该...吗”、“最好的方法是什么...”、“哪种方案...”时
在
```
architect
```
、
```
Plan
```
或
```
blueprint
```
工作流期间
做出非显而易见的权衡决策时

When to Skip

不适用场景

Do NOT apply when:

Executing already-decided implementation
Single obvious path exists (no real alternatives)
User explicitly chose the approach ("use X to do Y")
Task is mechanical/procedural, not decisional
Trivial choices with negligible impact

以下情况请勿应用本协议：

执行已确定的实施方案时
存在单一明确路径（无实际替代方案）时
用户明确指定了方案（“用X来做Y”）时
任务为机械/程序性工作，而非决策性工作时
影响可忽略的琐碎选择时

The Protocol

协议步骤

Step 1: Identify the Commitment

步骤1：明确决策内容

Before recommending an approach, explicitly state:

What decision is being made
What approach you're inclined toward
Why you're drawn to it

在推荐方案前，需明确说明：

正在做出什么决策
你倾向于哪种方案
你倾向该方案的原因

Step 2: Steel-Man the Opposition

步骤2：强化反方观点

Present the strongest case AGAINST your inclination:

What could go wrong?
What are you assuming that might be false?
What would a smart critic say?
What's the opportunity cost?
Under what conditions would this fail?

Requirements:

Be genuinely adversarial, not token objections
Attack the strongest version of your argument
Include at least one non-obvious failure mode

提出反对你倾向方案的最强理由：

可能会出现什么问题？
你做出了哪些可能错误的假设？
精明的批评者会怎么说？
机会成本是什么？
在什么条件下该方案会失败？

要求：

真正站在对抗角度，而非敷衍的反对意见
针对你的论点的最强版本进行反驳
至少包含一个非显而易见的失败模式

Step 3: Defend or Pivot

步骤3：辩护或转向

After the adversarial pass:

Explain why the approach might still be correct despite objections
What conditions make this the right choice?
What would need to be true for alternatives to win?
OR: Acknowledge the objections changed your recommendation

完成对抗性分析后：

解释为何尽管存在反对意见，该方案仍可能是正确的
什么条件下这是正确的选择？
什么情况下替代方案会更优？
或者：承认反对意见改变了你的推荐

Step 4: Present with Confidence Calibration

步骤4：给出带置信度校准的结论

Final recommendation should include:

Clear recommendation with reasoning
Key assumptions that must hold
Conditions that would invalidate this choice
Monitoring signals to watch for

最终推荐应包含：

清晰的推荐及理由
必须成立的关键假设
会使该选择失效的条件
需要关注的监控信号

Output Format

输出格式

markdown

undefined

markdown

undefined

Decision: [What's being decided]

Initial Inclination

[Approach] because [reasons]

Adversarial Challenge

Against this approach:

[Strong objection 1]
[Strong objection 2]
[Non-obvious failure mode]

What I might be wrong about:

[Assumption that could be false]

Against this approach:

[Strong objection 1]
[Strong objection 2]
[Non-obvious failure mode]

What I might be wrong about:

[Assumption that could be false]

Resolution

[Why it's still correct OR why I'm changing recommendation]

Recommendation: [Final choice]

Key assumptions: [What must be true]
Watch for: [Signals this was wrong]

undefined

Key assumptions: [What must be true]
Watch for: [Signals this was wrong]

undefined

Relationship to Other Tools

与其他工具的关系

reasoning-verifier: Post-hoc verification of completed reasoning
devils-advocate: Pre-commitment challenge before reasoning solidifies
Use both: devils-advocate during planning, reasoning-verifier after execution

reasoning-verifier: 对已完成的推理进行事后验证
devils-advocate: 在推理固化前进行承诺前的挑战
结合使用：在规划阶段使用devils-advocate，在执行后使用reasoning-verifier

Underlying Principle

核心原理

LLMs commit to answers early and rationalize backward. This protocol interrupts that pattern by forcing exploration of the solution space before commitment crystallizes.

大语言模型（LLMs）会过早给出答案并事后合理化。本协议通过在承诺固化前强制探索解决方案空间，打破这种模式。