Prerequisites
Environment Detection
PROJECT_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || echo "unknown")
mkdir -p memory/artifacts/experience-manager memory/retrospectives
Invocation Methods
Method 1: Artifact Passing(Recommended)
- Detect artifact:
memory/artifacts/*/experience-*.json
Method 2: Direct User Invocation
- Use
/experience-managercommand
Detailed Description: Refer to detailed-guide.md
Overview
experience-manager is an independent knowledge management skill, responsible for project experience knowledgeaccumulation, retrieval, retrospective, reflection, achieving a knowledge loop, helping AI avoid repeating mistakes and continuously improve.
Core Capabilities:
- Experience Accumulation: After task completion, synchronously accumulate at four levels
- Experience Retrieval: Before task start, intelligently retrieve historical experience
- Experience Retrospective: At milestone completion, keep what works, change what doesn't
- Experience Reflection: Verify effectiveness, extract wisdom
- Error Prevention: Detect risks, provide prevention suggestions
- Effect Tracking: Record application effects, continuously optimize
Key Value:
- Avoid repeating mistakes, quickly reuse successful patterns
- Continuously accumulate knowledge assets, form a self-evolving loop
Core Principles
Four-Level Accumulation Mechanism
| Level | File | Dimension | Priority | Purpose |
|---|---|---|---|---|
| Rule Layer | CLAUDE.md | What to do | Highest | Project conventions, strong constraints |
| Strategy Layer | AGENT.md | How to do it | High | AI execution strategy, decision rules |
| Knowledge Layer | MEMORY.md | What was learned | Medium | Technical knowledge, experience summary |
| History Layer | MCP Memory | What happened in the past | Low | Complete history, intelligent retrieval |
Retrieval Priority: Rule Layer > Strategy Layer > Knowledge Layer > History Layer
Conflict Handling: Higher priority prevails.
Continuous Improvement Principles
- Crown the good: Successful experiences upgraded to rules/strategies
- Correct the bad: Failed lessons update the knowledge base
- Dynamic Update: Outdated experiences marked and updated in time
Core Capabilities
Capability 1: Experience Accumulation (Save)
Trigger Timing: After task completion, invoked by goal-oriented
Execution Steps:
- Receive request artifact (
experience-save-request.json) - Generate retrospective report (
memory/retrospectives/{task-id}.md) - Four-level synchronous accumulation (CLAUDE.md, AGENT.md, MEMORY.md, MCP Memory)
- Create result artifact (
save-result-{timestamp}.json)
Example:
// 输入
{
"action": "save",
"task_id": "2026-03-24_task-keywords",
"lessons_learned": ["经验1", "经验2"],
"errors_fixed": [{"error": "...", "fix": "..."}]
}
// 输出:四层级同步更新 + 结果工件
Capability 2: Experience Retrieval (Retrieve)
Trigger Timing: Before task start, invoked by goal-oriented, pilot
Execution Steps:
- Receive request artifact (
experience-request.json) - Four-tier parallel retrieval (keyword matching + semantic retrieval)
- Sort by priority and deduplicate
- Generate prevention checklist and recommended patterns
- Create result artifact (
experience-result-{timestamp}.json)
Example:
## 📚 历史经验参考
### ⚠️ 错误预防(来自 CLAUDE.md)
- 预防措施1
- 预防措施2
### 🎯 可复用模式(来自 AGENT.md)
- 代码模式/流程模式
### 📖 技术知识(来自 MEMORY.md)
- 知识点1、知识点2
Capability 3: Experience Review
Trigger timing: When a milestone is completed, invoked by goal-oriented
Purpose: Keep the good, change the bad, continuously optimize
Execution steps:
- Read the original review file
- Analyze application effectiveness (number of applications, success rate)
- Evaluate gains and losses (keep the good, change the bad)
- Update the four-tier knowledge base
- Generate review report
Example:
## 复盘结果
### ✅ 保留好的(好则加冕)
1. **成功经验**
- 应用效果:5次,成功率100%
- 决策:写入 CLAUDE.md 强制规则
### ❌ 改掉坏的(错则改之)
1. **失败教训**
- 问题:文档命名不直观
- 改进:优化命名规范
Capability 4: Experience Reflection
Trigger timing: Experience may be outdated, tech stack upgrades, periodic reflection
Purpose: Verify effectiveness, distill wisdom
Execution steps:
- Identify experiences that need reflection (outdated/low success rate)
- Deeply analyze effectiveness (application statistics, failure cases)
- Distill wisdom (surface experience → deep wisdom)
- Update the four-tier knowledge base (mark old experience, save new experience)
Example:
## 经验反思 - 目录检查
### 有效性分析
- 应用次数:8次,成功率:87.5%
- 问题:未考虑版本差异
### 智慧提炼
**表层经验**:测试前检查目录
**深层智慧**:防御性思维 + 版本意识 + 避免过度防御
**更新决策**:区分版本环境,优化检查流程
Capability 5: Error Prevention
Trigger timing: Potential error risk detected
Execution steps:
- Detect risks (missing configuration, incorrect API usage, etc.)
- Retrieve similar historical errors
- Generate prevention suggestions and checklist
- Show to user for confirmation
Example:
## 💡 错误预防提醒
检测到风险:国际化配置缺失
预防措施:
- 检查 next.config.js 中的 i18n 配置
- 使用标准配置模板
历史错误:2026-03-20 类似错误
Capability 6: Effect Tracking
Trigger timing: When experience is applied
Execution steps:
- Record application scenario and context
- Evaluate application effectiveness (whether errors were prevented, time saved)
- Update experience statistics
- Continuously optimize the knowledge base
Example:
{
"experience_id": "2026-03-24_i18n_config",
"applied_count": 3,
"error_prevented": 2,
"time_saved": "50分钟"
}
Difference between review and reflection
| Dimension | Review | Reflect |
|---|---|---|
| Object | Completed task/milestone | Existing experience |
| Timing | Some time after completion | When experience may be outdated |
| Purpose | Keep the good, change the bad | Verify effectiveness, distill wisdom |
| Depth | Medium (focus on gains and losses) | Deep (focus on essence) |
Usage examples
Scenario 1: Complete process after task completion
# 1. goal-oriented 创建目标
用户:"优化文档结构"
# 2. 自动检索经验
experience-manager → 四层级检索 → 返回预防措施
# 3. 执行任务
应用经验或最佳实践
# 4. 里程碑复盘
experience-manager.review → 保留好的,改掉坏的
# 5. 整体完成
experience-manager.deep-review → 提炼智慧
# 6. 沉淀经验
四层级同步更新
Scenario 2: Retrieve historical experience
# 用户提出任务
用户:"修复英文页面404错误"
# goal-oriented 调用
experience-manager.retrieve(
keywords=["英文页面", "404", "路由"]
)
# 返回结果
- 错误预防:检查 i18n 配置
- 可复用模式:国际化路由配置模板
- 历史案例:2026-03-20 修复详情
Common pitfalls
- ❌ Only accumulating without reading → Experience must be retrieved before each task starts
- ❌ Only accumulating in a single location → Accumulate synchronously across four tiers to ensure knowledge is not lost
- ❌ Only accumulating without reviewing → Review is required when milestones and the overall project are completed
- ❌ Outdated experience not reflected on → Reflect periodically and promptly update outdated experience
- ❌ Reviews become a formality → In-depth analysis of gains and losses, develop actionable improvement plans
- ❌ Neglecting wisdom distillation → Distill methodologies and transferable patterns from experience
Artifact formats
Request artifacts:
{
"requesting_skill": "skill-name",
"action": "save|retrieve|review|reflect",
"task_keywords": ["关键词"],
// 其他参数...
}
Result artifacts:
{
"status": "success",
"timestamp": "2026-03-24T10:00:00Z",
"layers_updated": ["CLAUDE.md", "AGENT.md", "MEMORY.md", "MCP Memory"],
// 结果数据...
}
Detailed format: Refer to templates/ directory
Template files
templates/retrospective-template.md- Retrospective report templatetemplates/experience-request-template.json- Retrieval request templatetemplates/experience-retrieve-template.json- Retrieval result templatetemplates/review-request-template.json- Retrospective request templatetemplates/reflection-request-template.json- Reflection request template
Reference materials
Core concepts:
Methodology:
- Double Loop Learning - Chris Argyris
- Kaizen - Continuous Improvement
- The Five Whys - Root Cause Analysis
Detailed guide: Refer to detailed-guide.md for complete execution steps, examples, and best practices.
Skill collaboration
Called entry points:
- ✅
goal-oriented- Retrieve when creating goals, retrospect and reflect upon milestone/overall completion, consolidate after completion - ✅
pilot- Retrieve before scheduling, prevent when detecting risks - ✅ Other skills - Called as needed
Invocation method: Artifact transfer (recommended) or direct user invocation