🔄 Self-Iteration Engine#

Self-iteration and feedback learning engine for AI agent skills. Tracks usage logs, detects performance patterns, triggers skill updates, and proposes new skill creation based on repeated request patterns.

This is a shared component skill — other skills reference it for self-improvement. When updating, ensure backward compatibility with all dependent skills. Users may install this skill standalone for its capabilities.

Usage Log Format#

Maintain a usage log file for each skill that declares dependency:

markdown

# Usage Log: <skill-name>
## [YYYY-MM-DD]
- Request: <brief description>
- Outcome: success | partial | fail
- User corrected: yes | no
- Correction detail: <if yes, what was corrected>
- Lesson: <what to do differently next time>

File location: memory/usage-logs/<skill-name>.md

Self-Iteration Triggers#

Evaluate these conditions during each periodic review (default daily, configurable):

ConditionAction3+ consecutive successful invocationsMark skill as "stable" — reduce context allocation2+ failures for the same scenarioFlag for SKILL.md reassessmentSame request type appears 3+ timesEvaluate creating a new dedicated skillUser corrected outputLog correction, adjust future behavior for that scenarioSkill hasn't been reviewed in 30+ daysTrigger review: check if dependencies changed, update examplesExternal tech change detectedCompare against skill's core technology stack, update if needed

Feedback Loop Implementation#

yaml

# memory/feedback-loop/<skill-name>.yaml
feedback_loop:
  last_review: "2026-05-19"
  next_review: "2026-05-26"
  status: "stable" | "needs-attention" | "monitoring"
  patterns_observed:
    - pattern: "user asks for financial data on weekends"
      current_response: "check if markets are open"
      improvement: "pre-fill with last trading day data"
      status: "resolved" | "pending" | "in-progress"
  skill_performance:
    total_calls: 47
    success_rate: 0.96
    issues:
      - "data freshness on weekends"

Review Cycle#

Daily (lightweight)#

• Scan today's usage log entries
• Check for failure patterns
• Log any user corrections

Weekly (moderate)#

• Aggregate performance stats
• Check iteration triggers (listed above)
• If any trigger fires → update SKILL.md or create new skill
• Archive usage logs older than 7 days

Monthly (deep)#

• Full performance review across all skills
• Compare success rates, identify declining trends
• Check if any external technology replaced the skill's core stack
• Propose new skill ideas based on accumulated pattern data
• Run memory cleanup (delegate to complex-memory-manager)

Update Decision Matrix#

SignalDecision80%+ success rate, no user correctionsNo update needed60-80% success rateMinor update: clarify instructions, add edge cases<60% success rateMajor update: redesign workflow, check data sourcesUser corrects same thing 3+ timesFix that specific guidance in SKILL.mdExternal API / tool changedUpdate immediatelyNew competing technology availableEvaluate migration; update if 2x+ better

New Skill Creation Criteria#

Create a new skill when:

• Same request pattern appears 3+ times across different users
• The pattern cannot be handled well by existing skills
• The pattern has a clear, bounded scope
• A distinct tool/API would improve the result

Document in memory/skill-ideas/:

yaml

proposal:
  name: <suggested-slug>
  rationale: "Pattern X appeared N times. Existing skill Y handles it poorly because..."
  scope: "<bounded description>"
  priority: high | medium | low
  created: <date>

Cross-Skill Usage#

Other skills declare dependency:

yaml

metadata:
  openclaw:
    requires:
      skills:
        - self-iteration-engine

Usage logs are prefixed with the source skill name:

• memory/usage-logs/skill-a.md
• memory/feedback-loop/skill-a.yaml

When this skill updates log format, check ALL dependent skills' parsing logic.

---

🔄 自迭代引擎#

面向AI Agent技能的自迭代与反馈学习引擎。追踪使用日志、检测性能模式、触发技能更新，并基于重复请求模式提出新技能创建建议。

这是一个 共享组件技能——其他技能通过它实现自我改进。更新时需保证向后兼容。用户也可能独立安装此技能使用其能力。

使用日志格式#

每个声明依赖的技能维护一份使用日志：

markdown

# 使用日志：<skill名称>
## [YYYY-MM-DD]
- 请求：<简述>
- 结果：成功 | 部分成功 | 失败
- 用户修正：是 | 否
- 修正详情：<如果是，修正了什么>
- 经验：<下次应该怎么做>

文件位置： memory/usage-logs/<skill名称>.md

自迭代触发条件#

定期审查 时评估以下条件（默认每天，可配置）：

条件行动连续成功3次以上标记为"稳定"——减少上下文分配同一场景失败2次以上标记SKILL.md需重新评估同类请求出现3次以上评估创建新专用skill用户修正了输出记录修正，调整后续该场景的行为技能超过30天未审查触发审查：检查依赖是否变更、更新示例检测到外部技术变化与技能核心技术栈对比，需要时更新

反馈循环实现#

yaml

# memory/feedback-loop/<skill名称>.yaml
feedback_loop:
  last_review: "2026-05-19"
  next_review: "2026-05-26"
  status: "stable" | "needs-attention" | "monitoring"
  patterns_observed:
    - pattern: "用户在周末查询金融数据"
      current_response: "检查市场是否开盘"
      improvement: "自动填充最近交易日数据"
      status: "resolved" | "pending" | "in-progress"
  skill_performance:
    total_calls: 47
    success_rate: 0.96
    issues:
      - "周末数据新鲜度"

审查周期#

每日（轻量）#

• 扫描今天的日志条目
• 检查失败模式
• 记录用户修正

每周（中等）#

• 汇总性能统计
• 检查触发条件
• 触发更新或创建新技能
• 归档超过7天的日志

每月（深度）#

• 全技能性能审查
• 对比成功率，识别下降趋势
• 检查外部技术是否取代了技能核心
• 基于积累的模式数据提出新技能想法
• 执行记忆清理（委托给complex-memory-manager）

更新决策矩阵#

信号决策成功率>80%，无用户修正无需更新成功率60-80%小幅更新：澄清说明、补充边界情况成功率<60%重大更新：重新设计工作流、检查数据源用户修正同一内容3次以上在SKILL.md中修复该指导外部API/工具变更立即更新出现新的竞争技术评估迁移；若2倍以上优于现有则更新

新技能创建标准#

以下情况创建新技能：

• 相同请求模式在不同用户出现3次以上
• 现有技能无法良好处理该模式
• 该模式有清晰、有边界的范围
• 有独特工具/API可提升结果

记录在 memory/skill-ideas/：

yaml

proposal:
  name: <建议的slug>
  rationale: "模式X出现了N次。现有技能Y处理不佳因为..."
  scope: "<有边界的描述>"
  priority: high | medium | low
  created: <日期>

跨技能使用#

其他技能声明依赖的方式：

yaml

metadata:
  openclaw:
    requires:
      skills:
        - self-iteration-engine

使用日志以源技能名称为前缀：

• memory/usage-logs/skill-a.md
• memory/feedback-loop/skill-a.yaml

本技能更新日志格式时，需检查所有依赖技能的解析逻辑。