← 返回 Skills 市场
AI Growth Engine
作者
KingOfZhao
· GitHub ↗
· v1.0.0
· MIT-0
200
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install ai-growth-engine
功能描述
AI成长引擎 —— 通用自我迭代框架,回顾→提取模式→调参→验证→记录,任何Agent/职业都可用的成长操作系统
使用说明 (SKILL.md)
AI Growth Engine
元数据
| 字段 | 值 |
|---|---|
| 名称 | ai-growth-engine |
| 版本 | 1.0.0 |
| 作者 | KingOfZhao |
| 发布日期 | 2026-03-31 |
| 置信度 | 96% |
来源碰撞
programmer-cognition (代码成长)
⊗
researcher-cognition (科研成长)
⊗
self-evolution-cognition (通用自进化)
↓
ai-growth-engine (通用AI成长引擎)
核心哲学
程序员通过Code Review成长,科研人员通过假设验证成长,设计师通过视觉评审成长。 成长的方式不同,但成长的引擎相同。
ai-growth-engine就是这个引擎——职业无关、Agent通用。
RAPVL 五步成长循环
这是从programmer-cognition和researcher-cognition中提取的元模式:
R — Review 回顾: 审视最近的N次行动(代码提交/实验结果/设计迭代)
A — Analyze 提取模式: 从成功和失败中提取模式(bug模式/假设失败模式/设计缺陷模式)
P — Plan 调参: 基于模式制定改进计划(修改代码习惯/调整实验设计/更新设计流程)
V — Verify 验证: 执行改进计划并用度量指标验证(测试通过率/假设验证率/视觉一致性)
L — Learn 记录: 将验证结果写入成长日志,更新内部模型(永久记忆)
RAPVL vs 其他成长模型
Plan-Do-Check-Act (PDCA): 4步,缺少"模式提取"和"记忆持久化"
OODA Loop: 4步,面向决策而非成长
RAPVL (本Skill): 5步,专为AI自进化设计,包含模式提取+文件记忆
成长度量化(Growth Score™)
成长不能只靠感觉,必须有数字。
基础公式:
Growth Score = Σ (round_success_rate_delta × task_complexity_weight)
计算示例(程序员):
Round 1: 10个任务, 7个成功 (70%), 复杂度均值0.6
Round 2: 10个任务, 8个成功 (80%), 复杂度均值0.6
Growth Score = (0.80 - 0.70) × 0.6 = 0.06 (+6%)
计算示例(科研人员):
Round 1: 5个假设, 1个验证通过 (20%), 重要度均值0.8
Round 2: 5个假设, 2个验证通过 (40%), 重要度均值0.8
Growth Score = (0.40 - 0.20) × 0.8 = 0.16 (+16%)
成长趋势:
- Growth Score > 0 → 正在成长 ✅
- Growth Score = 0 → 停滞 ⚠️ 触发"深度反思"模式
- Growth Score \x3C 0 → 退步 🔴 触发"根因分析"模式
职业适配层(引擎相同,指标不同)
# 成长引擎是通用的,只需传入不同的度量配置:
growth_configs = {
"程序员": {
"success_metric": "测试通过率 + CI通过率",
"failure_patterns": ["空指针", "竞态条件", "API变更遗漏"],
"complexity_weight": "代码行数 × 依赖数 × 并发复杂度",
"verification": "单元测试 + 集成测试 + 生产监控",
"record": "CHANGELOG + debug日志 + LEARNINGS.md"
},
"科研人员": {
"success_metric": "假设验证通过率 + 复现成功率",
"failure_patterns": ["混淆变量未控制", "样本量不足", "选择性报告"],
"complexity_weight": "实验变量数 × 跨领域程度 × 数据规模",
"verification": "统计显著性 + 同行评审 + 可复现性检查",
"record": "hypotheses.md + experiments/ + collision_log/"
},
"设计师": {
"success_metric": "视觉一致性评分 + 用户满意度",
"failure_patterns": ["对齐偏移", "色彩不一致", "可用性问题"],
"complexity_weight": "页面数 × 交互复杂度 × 跨设备适配",
"verification": "设计系统检查 + A/B测试 + 用户测试",
"record": "design_log/ + iteration_history/"
},
"企业家": {
"success_metric": "关键指标增长率 + 决策准确率",
"failure_patterns": ["市场误判", "资源错配", "时机错误"],
"complexity_weight": "市场规模 × 竞争强度 × 不确定性",
"verification": "MVP验证 + 市场反馈 + 财务指标",
"record": "decision_log/ + pivot_history/"
},
"教师": {
"success_metric": "学生理解率 + 考试通过率",
"failure_patterns": ["概念跳跃", "练习不足", "反馈延迟"],
"complexity_weight": "班级人数 × 学科难度 × 基础差异",
"verification": "随堂测验 + 作业分析 + 期末评估",
"record": "teaching_log/ + student_progress/"
},
"医生": {
"success_metric": "诊断准确率 + 治疗有效率",
"failure_patterns": ["误诊", "过度治疗", "信息遗漏"],
"complexity_weight": "病例复杂度 × 症状模糊度 × 时间压力",
"verification": "随访结果 + 同行会诊 + 临床指南对照",
"record": "case_log/ + differential_diagnosis/"
}
}
自进化机制
成长引擎自身的进化(元进化):
1. 每完成一轮RAPVL循环,记录:
- Growth Score
- 哪种失败模式提取最有效
- 哪种调参策略效果最好
- 哪种验证方式最准确
2. 定期(每10轮)分析:
- 成长加速因子: 什么让Growth Score提升最快?
- 成长瓶颈: 什么导致Growth Score停滞?
- 优化方向: 调整模式提取策略或验证方式
3. 输出进化报告:
growth_reports/evolution_{round}.json
安装命令
clawhub install ai-growth-engine
# 或手动安装
cp -r skills/ai-growth-engine ~/.openclaw/skills/
调用方式
from skills.ai_growth_engine import AIGrowthEngine
# 通用初始化(任何Agent都可以用)
engine = AIGrowthEngine(workspace=".")
# 使用预设配置
engine.configure(profession="程序员")
# 或自定义配置
engine.configure_custom(
success_metric="自定义指标",
failure_patterns=["模式1", "模式2"],
complexity_weight_fn=lambda task: task.difficulty * task.dependencies
)
# 完成一轮RAPVL
result = engine.run_rapvl_round(
actions=[
{"task": "修复登录bug", "success": True, "complexity": 0.7},
{"task": "优化查询性能", "success": False, "complexity": 0.9},
{"task": "添加单元测试", "success": True, "complexity": 0.4},
]
)
print(result.growth_score) # +0.04
print(result.top_patterns) # ["API变更遗漏频率↑", "性能优化成功率低"]
print(result.recommendations) # ["增加API变更检测工具", "性能优化前先profile"]
print(result.trend) # "growing" / "stagnant" / "declining"
# 查看成长历史
history = engine.growth_history()
print(history.rounds) # [Round1, Round2, ...]
print(history.cumulative_growth) # 0.42 (+42%)
print(history.best_domain) # "代码审查"
print(history.worst_domain) # "性能优化"
# 强制深度反思(Growth Score = 0时自动触发)
reflection = engine.deep_reflection()
print(reflection.root_cause) # "过度追求新功能,忽略代码质量"
print(reflection.action_plan) # ["每3个功能后强制1次重构", ...]
学术参考文献
- A Survey of Self-Evolving Agents — 自进化综述(RAPVL的理论基础)
- SAGE: Multi-Agent Self-Evolution — 四Agent闭环验证
- Group-Evolving Agents — 群体进化+经验共享
- Self-evolving Embodied AI — 元进化+自改进循环
- Memory in the Age of AI Agents — 成长记忆持久化
- Beyond RAG for Agent Memory — 跨轮次经验聚合
安全使用建议
This package looks like documentation and an API spec rather than runnable code. Before installing or running it: 1) Confirm the actual code source — inspect the referenced GitHub repo (https://github.com/KingOfZhao/AGI_PROJECT) or the registry entry that 'clawhub install' would pull from; do not run an installer that downloads unknown archives. 2) If you intend to use it, review the implementation code for any network calls, telemetry, or unexpected file system operations. 3) Be aware the skill expects to write reports/history to your workspace — avoid using sensitive directories, or run it in an isolated workspace/container. 4) If you're uncertain about the provenance, treat this as documentation-only (implement the logic yourself) rather than blindly installing external code. Verifying the actual package contents and install URL would change this assessment to benign if they match the described API and are from a trusted source.
能力评估
Purpose & Capability
The SKILL.md advertises a concrete Python API (AIGrowthEngine class, methods like configure, run_rapvl_round, growth_history) and explicit outputs (growth_reports/*.json). However, the skill bundle contains only documentation files (SKILL.md, README, VERIFICATION_PROTOCOL.md, HEARTBEAT.md) and no code files or declared install steps. That means the claimed runtime capabilities are not present in this package — either the skill is purely an instruction/template or it expects an external package to be installed. This disconnect is disproportionate to the stated purpose and should be clarified.
Instruction Scope
The instructions themselves are domain-focused (RAPVL loop, metrics, file paths for reports and history) and don't ask the agent to access unrelated credentials or system-wide secrets. They do assume the ability to read/write workspace files and to call a Python module. Because no implementation is included, the SKILL.md may implicitly require installation of external code or creation of files; verify where that code would come from and whether it will write to your workspace or other paths.
Install Mechanism
No install spec is provided inside the skill bundle despite the README/SKILL.md showing an install command (clawhub install ai-growth-engine) and import examples. That suggests installation relies on an external registry or repository. Without an install specification or included code, it's unclear what 'clawhub install' would fetch (which URL, which release, what code). This gap increases risk because arbitrary code could be pulled at install time.
Credentials
The skill does not request any environment variables, credentials, or special config paths. The lack of requested secrets is proportionate to the stated purpose (tracking/metrics and local reports).
Persistence & Privilege
The skill describes persistent outputs (growth_reports, history files) and automated triggers (daily runs, every-10-round reports). always is false and autonomous invocation is allowed by default. This is not inherently problematic, but you should expect the skill (or the external code it installs) to read and write files in the agent workspace and possibly run periodic tasks; verify what will be persisted and whether those files could contain sensitive data.
如何使用
- 确保已安装 OpenClaw(本地或 Docker 部署)
- 在对话框中输入安装命令:
/install ai-growth-engine - 安装完成后,直接呼叫该 Skill 的名称或使用
/ai-growth-engine触发 - 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.0
Skill工厂: RAPVL五步成长循环+Growth Score量化+6职业预设+元进化
元数据
常见问题
AI Growth Engine 是什么?
AI成长引擎 —— 通用自我迭代框架,回顾→提取模式→调参→验证→记录,任何Agent/职业都可用的成长操作系统. 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 200 次。
如何安装 AI Growth Engine?
在 OpenClaw 或 Claude Code 对话框中运行命令「/install ai-growth-engine」即可一键安装,无需额外配置。
AI Growth Engine 是免费的吗?
是的,AI Growth Engine 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。
AI Growth Engine 支持哪些平台?
AI Growth Engine 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。
谁开发了 AI Growth Engine?
由 KingOfZhao(@kingofzhao)开发并维护,当前版本 v1.0.0。
推荐 Skills