session-intelligence-harvester

会话智慧收集器

概述

通过在 RII 组件中实施更新，将会话中的经验转化为永久的组织智慧。在涉及纠正、发现或模式识别的高效会话后，系统地提取所学内容，将其分配到正确的组件，并应用变更。

为何重要：未编码到 RII 组件中的一次性修复问题会反复出现。Chapter N 技能格式漂移之所以发生，是因为没有检查来防止它。收集经验后，这种故障模式已编码到 4 个文件中——未来的会话将自动受益。

何时使用此技能

自动触发（主动建议收集经验）：

• 会话纠正了格式漂移（文件结构、YAML、调用方式错误）
• 会话向编排文件添加了缺失的检查
• 会话识别出值得预防的故障模式
• 会话以相似的模式纠正了 3 个或更多文件
• 创建了记录重要经验的 PHR

手动触发（用户请求）：

• “收集本次会话的经验”
• “记录会话智慧”
• “我们应该从这项工作中总结出什么？”
• “根据我们的经验更新 RII”

工作流

默认执行操作：实施所有更新，而不只是提出建议。读取目标文件、进行编辑并提交变更。只有在明确提出要求时，才只给出建议而不实施。

第 1 步：会话分析

通过回答以下问题来分析会话。将分析写下来以跟踪进度：

1. CORRECTIONS MADE
   - What errors/drift were corrected?
   - What was wrong vs what is now correct?
   - Why did the error occur? (missing check, format drift, etc.)

   WHY THIS MATTERS: Understanding root cause determines which RII
   component prevents recurrence. Format drift → agent convergence
   pattern. Missing context → CLAUDE.md protocol step.

2. PATTERNS IDENTIFIED
   - What recurring patterns emerged?
   - What canonical sources were referenced?
   - What validation would have caught this earlier?

   WHY THIS MATTERS: Patterns that recur across sessions deserve
   encoding. If you referenced a canonical source, other sessions
   will need that same reference.

3. LEARNING CLASSIFICATION
   - Context-gathering gap? (CLAUDE.md)
   - Pedagogical/teaching issue? (Constitution)
   - Agent convergence pattern? (Agents)
   - Reusable workflow? (Skills)
   - Missing orchestration check? (Commands)

   WHY THIS MATTERS: Wrong routing means learnings don't trigger at
   the right time. A convergence pattern in CLAUDE.md won't help
   chapter-planner catch it during planning.

第 2 步：分配到 RII 组件

使用此路由表。将经验分配到能在恰当时机发现它们的组件：

经验类型	目标组件	位置	添加内容	触发时机
上下文收集缺口	CLAUDE.md	Section I	在上下文协议中添加新步骤	任何平台工作之前
故障模式示例	CLAUDE.md	Failure modes	带有纠正措施的命名示例	检测到类似情况时
教学框架	Constitution	Section IIa	教学方法更新	课程设计期间
Agent 收敛模式	Agent 文件	Convergence Patterns	模式、原因及纠正措施	Agent 执行期间
Agent 自我监控	Agent 文件	Self-Monitoring Checklist	新检查项	Agent 最终确定输出之前
规范来源查找	多个 Agent	Analysis Questions	交叉引用检查	规划阶段
可复用工作流	新技能	.claude/skills/	新的 SKILL.md	用户调用技能时
编排检查	命令文件	Phase 0 或相关阶段	新验证步骤	工作流执行期间
格式规范	规范来源章节	课程内容	权威格式	教授该模式时

路由为何重要：将经验放在错误的组件中，无法防止问题再次发生。如果错误发生在 chapter-planner 执行期间，在 content-implementer.md 中添加检查将无济于事。

第 3 步：读取目标文件并生成更新

对每条已识别的经验：

1. 读取目标文件，了解当前结构并找到准确的插入位置
2. 生成更新内容，并附上上下文以便准确插入
3. 跟踪进度，逐一处理多个文件

## Learning: [Brief Title]

**Type**: [Context gap | Failure mode | Convergence pattern | etc.]

**Target**: [File path]

**Current State** (after reading file):
[What's missing or incorrect - quote existing content if helpful]

**Exact Placement**:
[Which section, after which content - be specific enough to Edit]

**Content to Add**:
[Exact content, matching the file's style and format]

**Rationale**:
[Why this prevents recurrence - what will trigger this check]

**Canonical Source** (if applicable):
[Which chapter/lesson defines the authoritative format]

Step 4: 实施更新

采取行动：编辑每个目标文件。使用 Edit 工具进行修改。

每项更新：

1. 阅读目标文件（如果尚未阅读）
2. 找到确切的插入位置
3. 使用 Edit 工具应用修改
4. 验证修改已正确应用

跟踪完成情况：

Updates Progress:
- [x] CLAUDE.md - Added failure mode section
- [x] chapter-planner.md - Added convergence pattern 6
- [x] sp.loopflow.v2.md - Removed (deprecated, replaced by native Plan Mode)
- [ ] content-implementer.md - Pending

Step 5: 验证

完成前，逐项验证以下内容（确认后勾选）：

- [ ] Read each target file before editing (no speculation about structure)
- [ ] Verified each learning routes to the component where it triggers at the right time
- [ ] Confirmed updates include exact placement context (not vague locations)
- [ ] Checked canonical sources exist for format-related learnings
- [ ] Searched target files to confirm no duplicate information exists
- [ ] Used imperative form for agent files, appropriate style for others
- [ ] Added cross-references where pattern appears in multiple files
- [ ] All edits applied successfully (no pending changes)

Step 6: 创建 PHR 并提交

1. 创建 PHR，记录：

   • 学到了什么
   • 记录在哪些文件中（列出所有文件）
   • 为什么这能改进后续工作

2. 提交更改，并使用描述性消息：

   feat(intelligence): Harvest session learnings into RII
   
   Updates:
   - [File 1]: [What was added]
   - [File 2]: [What was added]
   
   Prevents: [What failure mode this prevents]
   

RII 组件参考

CLAUDE.md 结构

Section I: Context Gathering Protocol
  - Step 1-N: Sequential context steps
  - Each step has WHAT to do and WHY it matters
  - "Find canonical source" step for pattern teaching

Failure Modes (between Section I and II):
  - Named failure examples: "FAILURE MODE: [Name] Example"
  - "What I did wrong" list
  - "What I should have done" numbered steps
  - "Result" showing what was prevented

Agent 文件结构

Analysis Questions (Section III):
  - Numbered questions with "Why this matters" explanation
  - Self-check prompt at end

Principles (Section IV):
  - Named principles with Framework + What this means + Application guidance
  - Self-check prompt

Convergence Patterns (Section VI):
  - "Generic pattern" description
  - "Why this is convergence" explanation
  - "Correction" with specific steps

Self-Monitoring Checklist (Section VIII):
  - Numbered checklist with checkmark emoji prefix
  - Each item is a verification question

命令文件结构

Phase 0: Foundation checks
  - Constitutional reading
  - Canonical source checks (for educational content)
  - Each step explains WHY

Convergence Patterns:
  - Symptom description
  - Detection method
  - Correction steps

输出格式

完成整理后，提供摘要：

## Session Intelligence Harvest Summary

**Session**: [Brief description]
**Date**: [ISO date]
**Status**: COMPLETE

### Learnings Extracted: [N]

| #   | Learning | Type   | Target | Status  |
| --- | -------- | ------ | ------ | ------- |
| 1   | [Title]  | [Type] | [File] | Applied |
| 2   | [Title]  | [Type] | [File] | Applied |

### Updates Applied

1. **[File]**: [What was added] (lines X-Y)
2. **[File]**: [What was added] (lines X-Y)

### PHR Created

- Path: [PHR path]
- Stage: [Stage]

### Canonical Sources Referenced

- [Pattern]: [Chapter X Lesson Y]

### Commit

- Hash: [commit hash]
- Message: [commit message summary]

示例

示例 1：格式偏差修正（多文件）

会话：修正技能格式，使其采用基于领域的结构

分析：

CORRECTIONS MADE:
- Wrong: .claude/skills/section-writer.md (flat file, no domain)
- Correct: .claude/skills/authoring/section-writer/SKILL.md (domain + directory structure)
- Root cause: No domain organization for skills/agents

PATTERNS IDENTIFIED:
- Skills must be in authoring/ or engineering/ domain folders
- Agents must be in authoring/ or engineering/ domain folders
- Multiple files referenced old flat structure

CLASSIFICATION:
- Failure mode → CLAUDE.md
- Convergence pattern → chapter-planner.md, content-implementer.md
- Skill structure → skill-creator, session-intelligence-harvester

已应用的更新：

1. CLAUDE.md：更新 Agent 架构部分
2. skill-creator：添加领域组织要求
3. session-intelligence-harvester：更新路由表，加入领域路径
4. 将所有技能移至 authoring/ 或 engineering/
5. 将所有 Agent 移至 authoring/ 或 engineering/
6. 通用技能位于 .claude/skills/

示例 2：缺少验证（单文件）

会话：发现课程没有检查 chapter-index.md 中的先修要求

分析：

CORRECTIONS MADE:
- Wrong: Started chapter work without reading chapter-index.md
- Correct: MUST read chapter-index.md first to get Part, proficiency, prerequisites
- Root cause: No mandatory step in context protocol

CLASSIFICATION:
- Context-gathering gap → CLAUDE.md Section I

已应用的更新：

1. CLAUDE.md：添加 Step 1，要求阅读 chapter-index.md，并明确规定提取内容

示例 3：虚构事实（第 2 章事件）

会话：编写了 6 篇课程，其中包含未经验证的统计数据、日期和采用数量

分析：

CORRECTIONS MADE:
- Wrong: "50-75% time savings" for goose
- Correct: "75% of engineers save 8-10+ hours/week" (verified via Block announcement)

- Wrong: Conflated Agent Skills timeline (said single date)
- Correct: Oct 16, 2025 = Claude Code launch; Dec 18, 2025 = open standard release

- Wrong: Generic agent support lists
- Correct: Verified lists from official AAIF announcement

- Root cause: Trusted plausible-sounding data from memory instead of web verification

PATTERNS IDENTIFIED:
- Statistics, dates, and quotes MUST be web-verified before publication
- Existing factual-verifier agent was available but not used
- 50% of session time was spent fixing hallucinated facts

CLASSIFICATION:
- Critical failure mode → CLAUDE.md Failure Prevention + new section
- Process gap → Missing mandatory fact-check step

已应用的更新：

1. CLAUDE.md：添加“内容事实核查（强制）”部分
2. CLAUDE.md：在“失败预防”列表中添加失败模式示例
3. 记录 factual-verifier Agent 的调用模式

自我监控

在将整理标记为完成之前，确认你已：

• 分析会话，找出所有修正和模式
• 对每项经验进行分类，确定正确的路由
• 编辑前已阅读每个目标文件（不做臆测）
• 使用 Edit 工具应用所有修改（而不只是提出修改建议）
• 验证修改符合目标文件的风格和结构
• 在多个文件出现相同模式时添加交叉引用
• 创建 PHR，记录此次整理
• 使用描述性消息提交更改
• 生成摘要，展示已应用的所有更新