advanced-elicitation

高级引导

概述

将元认知推理应用于 AI 输出。通过 15 种以上的系统化方法，让 AI 重新审视自己的工作。

核心原则：初次回答通常不错，但还不够出色。引导能促使更深入的思考。

何时使用

适用于：

• 做出重要决策（架构、安全、重大功能）
• 解决复杂问题（涉及多个利益相关方、需求不明确）
• 产出关键成果（规范、计划、设计）
• 质量比速度更重要

不适用于：

• 简单查询（“X 是什么？”）
• 常规任务（格式调整、简单重构）
• 时间紧迫的情况（紧急修复）
• 预算受限的情况（成本增加 2 倍）

工作原理

1. 生成初始回答：Agent 给出初步答案
2. 应用引导方法：根据上下文选择 1-3 种方法
3. 重新审视：Agent 使用所选方法重新评估
4. 综合：整合见解，产出改进后的结果

引导方法

1. 第一性原理思维

描述：拆解至基本事实，从头重建推理过程

适用场景：

• 复杂系统设计
• 架构决策
• 创新挑战

提示词模板：

You are applying First Principles Thinking to:

---
{content}
---

Steps:
1. List all underlying assumptions
2. Question each assumption: "Is this fundamentally true?"
3. Identify fundamental truths (cannot be broken down further)
4. Rebuild solution from fundamentals only
5. Compare rebuilt solution to original - what changed?

Output:
### First Principles Analysis

**Fundamental Truths:**
- [Truth 1]
- [Truth 2]

**Assumptions Challenged:**
1. [Assumption] - [Why it might be wrong]

**Improvements:**
- [Improvement based on fundamentals]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

2. 事前验尸分析

描述：设想解决方案已经失败。逆向推演，找出原因。

适用场景：

• 规划重大变更
• 风险缓解
• 发布准备

提示词模板：

You are applying Pre-Mortem Analysis to:

---
{content}
---

Steps:
1. Fast-forward 6 months: the solution has failed spectacularly
2. List 5 reasons why it failed
3. For each reason, assess likelihood (Low/Medium/High)
4. For each high-likelihood failure, propose mitigation
5. Revise original solution with mitigations

Output:
### Pre-Mortem Analysis

**Failure Scenarios:**
1. [Scenario] - Likelihood: [L/M/H]
2. [Scenario] - Likelihood: [L/M/H]

**Mitigations:**
- [Mitigation for high-likelihood failures]

**Revised Solution:**
- [Changes to prevent failures]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

3. 苏格拉底式提问

描述：对每项假设都追问“为什么？”，直到找到根本依据。

适用场景：

• 需求分析
• 规范审查
• 消除歧义

提示词模板：

You are applying Socratic Questioning to:

---
{content}
---

Steps:
1. Identify 5 key claims in the content
2. For each claim, ask "Why is this true?"
3. For the answer, ask "Why?" again
4. Repeat until you hit a contradiction or fundamental truth
5. Revise claims that don't survive questioning

Output:
### Socratic Analysis

**Claim 1:** [Claim]
- Why? [Answer]
- Why? [Answer]
- Why? [Answer]
- **Verdict:** [Survives/Needs revision]

**Improvements:**
- [Changes after questioning]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

4. 红队对蓝队

描述：攻击解决方案（红队），为其辩护（蓝队），并综合得出改进方案。

适用场景：

• 安全审查
• 风险评估
• 对抗性测试

提示词模板：

You are applying Red Team vs Blue Team to:

---
{content}
---

Steps:
1. **Red Team**: List 5 ways to attack/break this solution
2. **Blue Team**: For each attack, propose a defense
3. **Red Team**: For each defense, find the weakness
4. **Blue Team**: Strengthen defenses
5. Synthesize: What changes make the solution more robust?

Output:
### Red Team vs Blue Team

**Attack 1:** [How to break it]
- Defense: [Blue team response]
- Counter-attack: [Red team finds weakness]
- Final defense: [Blue team strengthens]

**Improvements:**
- [Robust changes from adversarial testing]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

5. 反向思考

描述：不问“如何成功？”，而是问“如何失败？”，并避免这些情况。

适用场景：

• 风险识别
• 避免常见陷阱
• 负向空间分析

提示词模板：

You are applying Inversion to:

---
{content}
---

Steps:
1. Invert the goal: "How could we make this FAIL?"
2. List 5 ways to guarantee failure
3. For each failure mode, identify the opposite (success mode)
4. Check if original solution addresses success modes
5. Revise to explicitly avoid failure modes

Output:
### Inversion Analysis

**How to Fail:**
1. [Failure mode]
2. [Failure mode]

**How to Succeed (inverses):**
1. [Success mode]

**Improvements:**
- [Changes to avoid failures]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

6. 二阶思维

描述：考虑后果所带来的后果，以及长期影响。

适用场景：

• 战略决策
• 长期规划
• 权衡分析

提示词模板：

You are applying Second-Order Thinking to:

---
{content}
---

Steps:
1. Identify immediate consequences (1st order)
2. For each consequence, identify follow-on effects (2nd order)
3. For each 2nd order effect, identify further effects (3rd order)
4. Assess whether long-term effects align with goals
5. Revise solution to optimize for 2nd/3rd order effects

Output:
### Second-Order Analysis

**1st Order:** [Immediate effect]
- **2nd Order:** [Consequence of consequence]
  - **3rd Order:** [Further consequence]

**Long-Term Implications:**
- [Good/Bad long-term effects]

**Improvements:**
- [Changes optimizing for long-term]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

7. SWOT 分析

描述：优势、劣势、机遇、威胁。

适用场景：

• 战略规划
• 竞争分析
• 决策

提示词模板：

You are applying SWOT Analysis to:

---
{content}
---

Steps:
1. **Strengths**: What are the advantages?
2. **Weaknesses**: What are the disadvantages?
3. **Opportunities**: What external factors could help?
4. **Threats**: What external factors could harm?
5. Synthesize: How to leverage S+O, mitigate W+T?

Output:
### SWOT Analysis

**Strengths:**
- [Internal advantage]

**Weaknesses:**
- [Internal disadvantage]

**Opportunities:**
- [External positive factor]

**Threats:**
- [External negative factor]

**Strategy:**
- [Leverage strengths/opportunities, mitigate weaknesses/threats]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

8. 机会成本分析

描述：我们没有做什么？我们放弃了什么？

适用场景：

• 优先级排序
• 资源分配
• 权衡决策

提示词模板：

You are applying Opportunity Cost to:

---
{content}
---

Steps:
1. List what this solution requires (time, money, people)
2. List 3 alternative uses for those resources
3. For each alternative, estimate value
4. Compare: Is this solution the highest-value use?
5. If not, propose reallocation

Output:
### Opportunity Cost Analysis

**Resources Required:**
- [Time/Money/People]

**Alternatives:**
1. [Alternative use] - Estimated value: [X]
2. [Alternative use] - Estimated value: [Y]

**Verdict:**
- [Is this the best use? Why/why not?]

**Improvements:**
- [Reallocations or justifications]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

9. 类比推理

描述：其他人如何解决类似问题？从类比中汲取经验。

适用场景：

• 创新
• 从历史中学习
• 跨领域见解

提示词模板：

You are applying Analogical Reasoning to:

---
{content}
---

Steps:
1. Identify the core problem (abstract it)
2. Find 3 analogous situations (other domains/times)
3. How was the analogous problem solved?
4. What lessons transfer to this situation?
5. Adapt the solution based on analogies

Output:
### Analogical Analysis

**Core Problem:** [Abstract problem statement]

**Analogy 1:** [Domain/situation]
- How they solved it: [Solution]
- Lesson: [What transfers]

**Improvements:**
- [Adapted solution from analogies]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

10. 放宽约束

描述：如果约束 X 不存在，情况会怎样？解决方案会如何变化？

适用场景：

• 创新
• 打破假设
• 寻找创造性解决方案

提示词模板：

You are applying Constraint Relaxation to:

---
{content}
---

Steps:
1. List all constraints (explicit and implicit)
2. For each constraint, ask: "What if this wasn't true?"
3. Design solution without that constraint
4. Assess: Can we actually relax this constraint?
5. If yes, propose new solution. If no, learn from the thought experiment.

Output:
### Constraint Relaxation

**Constraint:** [Constraint]
- **If removed:** [Solution without constraint]
- **Can we actually relax it?** [Yes/No + reasoning]

**Improvements:**
- [Creative solutions from relaxation]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

11. 故障模式与影响分析（FMEA）

描述：可能出现什么问题？发生概率多大？后果有多严重？确定修复优先级。

适用场景：

• 工程设计
• 风险评估
• 安全关键系统

提示词模板：

You are applying FMEA to:

---
{content}
---

Steps:
1. List all components/steps in the solution
2. For each, identify potential failure modes
3. Rate each: Severity (1-10), Likelihood (1-10)
4. Calculate Risk Priority Number (RPN = Severity × Likelihood)
5. Address high-RPN failures first

Output:
### FMEA

**Failure Mode 1:** [What fails]
- Severity: [1-10]
- Likelihood: [1-10]
- RPN: [Product]
- Mitigation: [How to prevent/detect/recover]

**Improvements:**
- [Prioritized mitigations for high-RPN failures]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

12. 偏差检查

描述：哪些认知偏差可能影响判断？对其进行修正。

适用场景：

• 决策
• 审查流程
• 自我批评

提示词模板：

You are applying Bias Check to:

---
{content}
---

Steps:
1. Review common cognitive biases (confirmation, anchoring, sunk cost, availability, etc.)
2. For each bias, ask: "Is this affecting my reasoning?"
3. Find evidence of bias in the original content
4. Correct for identified biases
5. Re-evaluate the solution bias-free

Output:
### Bias Check

**Bias Detected:** [Bias name]
- **Evidence:** [Where it appears]
- **Correction:** [Adjusted reasoning]

**Improvements:**
- [Bias-free solution]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

13. 基准率思维

描述：类似情况下通常会发生什么？我们是否过于自信？

适用场景：

• 估算
• 风险评估
• 对乐观预期进行现实检验

提示词模板：

You are applying Base Rate Thinking to:

---
{content}
---

Steps:
1. Identify the reference class (similar past situations)
2. What's the base rate (average outcome for reference class)?
3. Why might this case be different?
4. Adjust estimates toward base rate (Bayesian update)
5. Revise solution with realistic expectations

Output:
### Base Rate Analysis

**Reference Class:** [Similar situations]
- **Base Rate:** [Typical outcome]
- **Our Estimate:** [Original estimate]
- **Adjusted Estimate:** [Reality-checked estimate]

**Improvements:**
- [More realistic solution]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

14. 钢人论证

描述：反方观点最有力的版本是什么？回应这个版本，而不是稻草人论证。

适用场景：

• 方案评审
• 辩论准备
• 坚持求知上的诚实

提示词模板：

You are applying Steelmanning to:

---
{content}
---

Steps:
1. Identify the opposing view (or alternative approach)
2. Strengthen it: What's the BEST argument against your solution?
3. Address the strong version (not a weak strawman)
4. If the steelman wins, adopt that approach
5. If your solution survives, it's stronger

Output:
### Steelman Analysis

**Opposing View:** [Alternative]
- **Strongest Argument:** [Best case for alternative]
- **Response:** [Addressing the strong version]
- **Verdict:** [Which approach is better?]

**Improvements:**
- [Refined solution after facing steelman]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

15. 切换时间跨度

描述：从 1 小时、1 天、1 个月、1 年和 5 年的时间跨度来看，这会是什么样子？

适用场景：

• 长期规划
• 权衡分析
• 策略评估

提示词模板：

You are applying Time Horizon Shift to:

---
{content}
---

Steps:
1. Evaluate solution at 1 hour: [Impact]
2. Evaluate at 1 day: [Impact]
3. Evaluate at 1 month: [Impact]
4. Evaluate at 1 year: [Impact]
5. Evaluate at 5 years: [Impact]
6. Identify time-horizon-dependent trade-offs
7. Optimize for the right time horizon

Output:
### Time Horizon Analysis

**1 Hour:** [Short-term effect]
**1 Day:** [Effect]
**1 Month:** [Effect]
**1 Year:** [Effect]
**5 Years:** [Long-term effect]

**Trade-Offs:**
- [Short vs long-term conflicts]

**Improvements:**
- [Optimized for appropriate horizon]

**Confidence Level:** [HIGH/MEDIUM/LOW]

---

使用模式

模式 1：单一方法（快速）

Skill({ skill: 'advanced-elicitation', args: 'first-principles' });

模式 2：多种方法（全面）

Skill({ skill: 'advanced-elicitation', args: 'first-principles,pre-mortem,red-team-blue-team' });

模式 3：自动选择（推荐）

Skill({ skill: 'advanced-elicitation', args: 'auto' });
// Automatically picks 2-3 methods based on content analysis

与 spec-critique 集成

高级启发可以增强 spec-critique：

// After generating spec
Skill({ skill: 'spec-critique', args: 'with-elicitation' });
// Applies elicitation to critique process

成本控制（根据 ADR-053）

• 仅限主动选择：绝不自动应用
• 预算上限：可通过 ELICITATION_BUDGET_LIMIT 配置
• 成本跟踪：与 cost-tracking hook 集成

配置：

features:
  advancedElicitation:
    enabled: true
    costBudget: 10.0 # USD per session
    minConfidence: 0.7 # Skip if confidence high
    maxMethodsPerInvocation: 5 # SEC-AE-001
    maxInvocationsPerSession: 10 # SEC-AE-003

安全控制

SEC-AE-001：输入验证

• 方法名称必须符合 /^[a-z][a-z0-9-]*$/
• 每次调用最多使用 5 种方法
• 无效方法将被拒绝并返回错误

SEC-AE-002：成本预算执行

• 启发前检查会话预算
• 跟踪累计成本
• 超出预算时优雅失败

SEC-AE-003：速率限制

• 每个会话最多进行 10 次启发
• 防止启发循环失控
• 超出限制时显示清晰的错误消息

示例

示例 1：架构决策

启发前：

We should use microservices with 12 services communicating via REST.

运用第一性原理后：

Fundamental truths: Services must communicate, data must be consistent.
Challenged assumption: "12 services" - is this the right granularity?
Could 6 bounded contexts suffice?

Improvement: Consolidate to 6-8 services by bounded context.
Use gRPC internally (40% latency reduction vs REST).

示例 2：安全评审

启发前：

JWT tokens for authentication across services.

红队／蓝队分析后：

Red Team Attack: Token theft via XSS, JWT validation on every call (latency).
Blue Team Defense: HttpOnly cookies, service mesh mTLS instead of JWT propagation.

Improvement: Use service mesh (Istio) for security instead of JWT propagation.

示例 3：规范验证

启发前：

Feature: User can delete their account.

事前验尸分析后：

Failure Scenario: 6 months later, GDPR compliance audit fails.
Cause: Deletion didn't cascade to all systems (analytics, backups).

Improvement: Add "Data Retention Audit" requirement.
Specify cascade delete to all systems within 30 days.

性能

• 质量提升：+30%（根据关键决策测得）
• 成本：LLM 使用量增加至 2 倍
• 耗时：+50%（对重要工作而言值得）

记忆协议（强制）

开始前：

cat .claude/context/memory/learnings.md

完成后：

• 新模式 → .claude/context/memory/learnings.md
• 发现问题 → .claude/context/memory/issues.md
• 作出决策 → .claude/context/memory/decisions.md

假设工作会被中断：如果没有记录在记忆中，就等于没发生。

铁律

1. 绝不自动应用启发——始终由用户主动选择。没有用户的明确请求或清晰的 Agent 意图信号，绝不调用。
2. 每种方法的输出都必须给出置信度等级——HIGH / MEDIUM / LOW 为必填项。未经校准的输出无法用于行动。
3. 每次调用绝不超过 5 种方法（SEC-AE-001）——过度启发只会产生噪声，而非信号。选择最相关的 1 至 3 种方法。
4. 调用前始终检查会话预算——超出 ELICITATION_BUDGET_LIMIT 时，应优雅失败并显示清晰的消息（SEC-AE-002）。
5. 绝不将启发视为证据的替代品——它能完善推理，但不能产生事实。结论始终应以代码库中的证据为依据。

反模式

反模式	失败原因	正确做法
自动应用于每个响应	简单任务的成本翻倍，却毫无收益	仅对重要或复杂的决策选择启用
一次运行全部 15 种方法	收益递减，令牌数量激增	选择 1–3 种最相关的方法
跳过置信度评级	未校准的评估毫无用处	始终输出 **Confidence Level:** HIGH/MEDIUM/LOW
用引导式提问取代证据	没有事实依据的推理只是猜测	先结合有依据的代码库证据，再进行引导式提问
不检查预算	会话成本不断攀升，却未被察觉	调用前始终检查 ELICITATION_BUDGET_LIMIT
在截止时间已到或紧急情况下运行	成本高昂，却没有时间落实改进	时间紧迫的修复应跳过；用于战略性工作

相关技能

• spec-critique - 规范验证（可调用引导式提问）
• security-architect - 安全审查（可使用引导式提问方法）
• verification-before-completion - 完成前检查

指定的 Agent

以下 Agent 可以使用此技能：

• planner - 用于战略决策
• architect - 用于架构审查
• security-architect - 用于威胁建模
• developer - 用于复杂的技术决策
• pm - 用于产品战略

---

版本：1.0.0 状态：生产就绪 作者：developer agent（任务 #6） 日期：2026-01-28