← Back to Skills Marketplace
joelee09

developer-self-improve-core

by Joe.Lee · GitHub ↗ · v1.1.9 · MIT-0
cross-platform ⚠ suspicious
164
Downloads
1
Stars
1
Active Installs
12
Versions
Install in OpenClaw
/install developer-self-improve-core
Description
开发者自改进核心技能 - 自动错误防重、自检、规则生成、记忆清洗、定时提醒 核心功能: - 每轮回答前:自动错误防重 - 每轮回答后:自动自检 + 生成规则草案 - 累计 10 轮对话/每周:自动记忆清洗扫描 - 自动提醒:每天 9:30 钉钉推送待确认规则 核心原则: - AI 只提议,人类终审 - 绝不自动写...
README (SKILL.md)

🛡️ developer-self-improve-core

开发者自改进核心技能


🚀 快速开始

1. 安装技能

clawhub install developer-self-improve-core

2. 初始测试(推荐)

首次安装后,建议先禁用自动化功能进行测试:

vi config/config.yaml

确认以下配置(默认已禁用):

enable_reminder: false      # 默认:false(测试时禁用)
enable_auto_cleanup: false  # 默认:false(测试时禁用)

测试命令:

./scripts/developer-self-improve-core.sh init
./scripts/developer-self-improve-core.sh pre-check "测试场景"
./scripts/developer-self-improve-core.sh post-check "测试内容" "测试"

3. 启用定时提醒(可选)

测试确认无误后,可以启用定时提醒:

  1. 配置钉钉账号:

    vi config/config.yaml
    
    dingtalk_target: "您的钉钉账号 ID"  # 替换为 18 位钉钉账号 ID
    enable_reminder: true               # 启用提醒
    
  2. 配置 crontab:

    crontab -e
    # 添加以下行(每天 9:30 执行):
    30 9 * * * cd ~/.openclaw/workspace/skills/developer-self-improve-core && ./scripts/daily-check.sh
    

4. 使用技能

技能会自动在 AI 回答前后执行:

  • ✅ 回答前:错误防重
  • ✅ 回答后:自检 + 规则生成

⚡ 触发时机

时机 执行内容 频率
每轮回答前 自动错误防重 每次响应前
每轮回答后 自动自检 + 生成规则草案 每次响应后
定期清洗 自动记忆清洗扫描 累计 10 轮对话或每周

🔄 一体化执行逻辑

1. 回答前(错误防重)

1. 加载当前场景对应的人工确认长期规则
   ↓
2. 命中已知错误类型则自动修正
   ↓
3. 仅使用人工确认规则,绝不使用未确认草案
   ↓
4. 内部标记:已规避同类错误

示例:

场景:代码推送
命中规则:【auto_push_001】永远使用 push_to_target.sh 脚本推送
行动:自动修正为使用脚本,标记"已规避同类错误"

2. 回答后(自检 + 反思 + 提案)

1. 静默自检:
   - 是否冲突长期规则
   - 是否重复犯错
   - 是否存在硬错误
   ↓
2. 仅在满足可信依据时生成规则草案:
   ✔ 用户明确规范/错误
   ✔ 可复现模式≥2 次
   ✔ 可验证逻辑/格式/事实错误
   ✔ 与历史规则冲突
   ↓
3. 禁止主观推断、特例泛化、幻觉规则
   ↓
4. 按固定格式生成草案,自动去重,不自动入库
   ↓
5. 主动询问用户:
   "我发现一个可沉淀的规则,是否加入长期记忆?【同意/修改/忽略】"
   ↓
6. 忽略的草案不再重复提示

可信依据验证:

依据类型 可信度 示例
用户明确要求≥2 次 "记住,永远用脚本推送"(第 2 次提及)
可复现模式≥2 次 域名切换问题出现 2 次
单次可验证硬错误 明显的逻辑/格式错误

3. 定期清洗(合并/去重/淘汰)

1. 自动扫描长期规则库:
   - 重复规则
   - 冲突规则
   - 低频规则(30 天未使用)
   - 被高阶规则覆盖的规则
   ↓
2. 生成清洗清单并询问:
   "以下规则建议合并/删除,是否执行?"
   ↓
3. 仅在用户批准后执行
   ↓
4. 保留操作日志,支持回滚

📝 强制格式约束

规则草案统一格式

### 【规则 ID】auto_xxxx
(基于场景哈希/内容指纹生成,自动去重)

### 【场景】xxx
(明确描述适用场景,≤20 字)

### 【问题/模式】xxx
(明确描述发现的问题或可复现模式,≤20 字)

### 【建议规则】xxx
(极简一句话,不冗余、可直接执行)

### 【依据】xxx
(用户明确指出/重复模式/可验证错误)

### 【可信度】高/中/低
(严格绑定依据,不可随意标注)

### 【建议层级】长期/中期

可信度绑定规则

可信度 判定标准
用户明确要求≥2 次
可复现模式≥2 次
单次可验证硬错误

⚙️ 安全铁律(不可突破)

铁律 说明
AI 只提议,人类终审 AI 可生成草案,但绝不自动入库
绝不自动写入/修改/删除长期记忆 所有记忆修改需用户批准
优先级排序 用户指令 > 长期规则 > AI 临时草案
记忆可追溯 所有记忆可追溯、可关闭、可回滚
规则冲突裁决 最新确认规则优先于旧规则
Token 优化 长期规则保持极简,仅一句话条目

📊 记忆结构

memory/
├── short_term/         短期记忆(对话结束自动销毁)
├── proposals/          临时提案区(按日期合并,保留 7 天)
│   ├── 2026-04-09.md   # 当天所有规则
│   └── 2026-04-08.md
├── rules/              长期规则库(仅人工确认)
│   └── confirmed_rules.md
├── cleanup/            待清洗区(需用户批准)
│   └── cleanup_20260409.md
└── logs/               操作日志(支持回滚)
    └── operations.log

🚀 使用示例

示例 1: 用户明确指出错误

对话:

用户:推送应该用 push_to_target.sh 脚本,不要手动 rsync
AI: 好的,已记录

AI 生成草案:

### 【规则 ID】auto_push_script_001
### 【场景】代码推送
### 【问题/模式】手动 rsync 推送产生重复文件
### 【建议规则】永远使用 push_to_target.sh 脚本推送
### 【依据】用户明确指出
### 【可信度】高
### 【建议层级】长期

AI 提案:

我发现一个可沉淀的规则,是否加入长期记忆?

【规则 ID】auto_push_script_001
【场景】代码推送
【建议规则】永远使用 push_to_target.sh 脚本推送

【同意/修改/忽略】

示例 2: 重复模式检测

对话:

用户:(第 1 次)域名切换后请求还是旧域名
AI: 已修复,添加了内存缓存

用户:(第 2 次)域名切换后请求还是旧域名
AI: 已修复,verifiedDomain 没有同步更新

AI 生成草案:

### 【规则 ID】auto_domain_cache_001
### 【场景】域名切换
### 【问题/模式】域名切换后 verifiedDomain 未同步更新
### 【建议规则】域名切换时必须同步更新 verifiedDomain 和 verifiedDomainCache
### 【依据】重复模式≥2 次
### 【可信度】中
### 【建议层级】长期

示例 3: 定期清洗

AI 提案:

累计 10 轮对话,触发记忆清洗扫描。

发现以下规则建议合并/删除,是否执行?

## 待合并规则

1. 【auto_push_001】永远使用 push_to_target.sh 脚本推送
2. 【auto_push_002】推送前检查配置文件
   ↓ 建议合并为
   【auto_push_001】推送相关规范(包含脚本使用和配置检查)

## 待删除规则

1. 【auto_old_rule_002】过时的规则(已 30 天未使用)
   ↓ 建议删除

【同意/修改/忽略】

🔧 脚本命令

初始化

./scripts/developer-self-improve-core.sh init

回答前检查(错误防重)

./scripts/developer-self-improve-core.sh pre-check "当前场景"

回答后检查(自检 + 提案)

./scripts/developer-self-improve-core.sh post-check "对话内容"

定期清洗

./scripts/developer-self-improve-core.sh cleanup

确认规则

./scripts/developer-self-improve-core.sh confirm [规则 ID] [同意/修改/忽略]

📖 更多文档

详见技能目录中的文档文件。


🤝 贡献

欢迎提交 Issue 和 Pull Request!

作者: lijiujiu
许可证: MIT


📄 许可证

MIT License

Usage Guidance
This skill appears to implement the claimed features, but take these precautions before enabling automation: 1) Reconcile metadata: confirm whether the skill actually requires AUTO_MEMORY_WORKSPACE and workspace/config/current_user.json (SKILL.md and .clawhub.json say yes, the top-level summary said no). 2) Inspect config/config.yaml and change dingtalk_target from the placeholder before enabling reminders; understand that enabling reminders will post proposal contents to DingTalk via your openclaw CLI account. 3) Keep enable_auto_cleanup and enable_reminder set to false while you test: run ./scripts/developer-self-improve-core.sh init, pre-check and post-check manually and inspect memory/ and logs/ to verify behavior. 4) Back up the skill directory before enabling automation (SECURITY.md recommends this). 5) If you plan to enable automated cleanup or cron jobs, review scripts cleanup_expired_proposals() and daily-check.sh to ensure they only act on the intended files — the cleanup function deletes expired proposal files without per-file confirmation. 6) Confirm openclaw CLI credentials are correctly configured and understand where notifications will be sent. If you need help auditing specific lines of the scripts (for example, the full cleanup/counter logic that was truncated), provide the full script and I can point out exact lines to change to enforce an approval step before deletions or external sends.
Capability Analysis
Type: OpenClaw Skill Name: developer-self-improve-core Version: 1.1.9 The skill is a developer productivity tool designed to help AI agents track coding rules and avoid repeating errors. It implements a robust 'human-in-the-loop' architecture where the AI can only propose rule changes, requiring explicit user confirmation before modifying long-term memory. Security features include a detailed SECURITY.md, operational logging in `memory/logs/operations.log`, and a default-off configuration for automated features like DingTalk reminders and crontab-based daily checks. The scripts (`developer-self-improve-core.sh`, `daily-check.sh`) use standard system utilities and the OpenClaw CLI for their stated purposes without evidence of obfuscation or malicious intent.
Capability Assessment
Purpose & Capability
Functionality (pre/post checks, proposals, memory cleanup, optional DingTalk reminders) aligns with the name/description. The scripts legitimately read their own config and an optional workspace/current_user.json to tag proposals. However, registry-level metadata shown at the top of the report lists no required env/config, while SKILL.md and .clawhub.json declare required binaries and an env (AUTO_MEMORY_WORKSPACE) and a config path (workspace/config/current_user.json) — that metadata mismatch is inconsistent and should be reconciled.
Instruction Scope
Runtime instructions and scripts read local config/config.yaml, optional workspace/config/current_user.json, write proposal/rule/log files under memory/, and (if enabled) send proposal summaries via the openclaw CLI to DingTalk. Those are in-scope for this skill, but the cleanup logic will delete proposal files older than retention_days when cleanup() / cleanup_expired_proposals() runs — this deletion is unconditional in the function and does not prompt for per-file human approval, which slightly contradicts the stated 'AI only proposes / human final approval' safety principle unless auto-cleanup remains disabled. Also daily-check.sh will assemble proposal contents and send them externally (DingTalk) via openclaw; enabling reminders will transmit proposal data outside the host.
Install Mechanism
No remote downloads or installers; this is an instruction+script package delivered in cleartext. No external install URLs or archive extracts were used. Risk from install mechanism itself is low.
Credentials
The skill does not request API keys or broad cloud credentials; it uses a local config value dingtalk_target (user-supplied DingTalk ID) and optionally relies on the openclaw CLI for notifications (openclaw manages its own auth). It reads workspace/config/current_user.json for user/platform tagging — plausible for its stated purpose but may expose identifying project/user fields. The earlier summary claiming 'Required env vars: none' conflicts with SKILL.md and .clawhub.json which list AUTO_MEMORY_WORKSPACE and various required binaries; verify which is authoritative before install.
Persistence & Privilege
The skill writes to its own memory/ directory and to logs (expected). always is false. However, it includes functions that can delete proposal files (cleanup_expired_proposals) without prompting when cleanup() runs, and cron/daily-check can send proposal contents externally. These behaviors are controlled by config flags (enable_auto_cleanup, enable_reminder) which default to disabled, but enabling automation gives the skill persistent behavior that can remove local files and send data externally — so only enable automation after testing and backups.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install developer-self-improve-core
  3. After installation, invoke the skill by name or use /developer-self-improve-core
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.1.9
Version 1.1.9 - 新增 .clawhub.json 和 .clawhub/origin.json 配置文件,支持 clawhub 平台管理。 - SKILL.md 文档优化,增加推荐的初始测试流程与配置项说明(enable_reminder/enable_auto_cleanup)。 - 更新钉钉提醒与定期自动化相关文档,明确测试推荐禁用自动化,测试无误后再开启。 - 移除对 jq 可选依赖,修改部分 requires 字段内容。 - 无核心功能变动,仅相关辅助文档和配置适配升级。
v1.1.8
v1.1.8 重新发布:确保脚本文件完整上传(developer-self-improve-core.sh 634 行,20KB)
v1.1.7
v1.1.7 安全优化: - 添加 env 和 config 声明(AUTO_MEMORY_WORKSPACE, workspace/config/current_user.json) - 完善 SECURITY.md(OpenClaw 凭证说明、自动清理说明、集成点说明、沙盒测试指导) - 更新 SKILL.md 元数据 - 清理多余文件
v1.1.6
v1.1.6 安全优化:setup-automation.sh 添加严格的 dingtalk_target 占位符检查,支持多种占位符格式检测
v1.1.5
v1.1.5 安全优化:移除硬编码钉钉 ID;添加 dingtalk_target 配置验证;修改默认 workspace 路径为当前目录;日志文件改为 skill 目录内;更新 SECURITY.md 权限说明
v1.1.4
v1.1.4 修复依赖声明:添加 md5sum, stat, sed 到必需依赖;添加 jq, openclaw 为可选依赖;完善 SECURITY.md 说明
v1.1.3
v1.1.3 安全优化:移除自动 crontab 操作,改为手动配置指导;完善 SECURITY.md 外部集成说明;更新依赖声明;memory 文件夹不上传(用户安装后自动创建)
v1.1.2
- Version bump to 1.1.2 with updated metadata. - Documentation updated in SKILL.md. - Removed unused memory/dialogue_counter.txt and memory/rules/confirmed_rules.md files.
v1.1.1
- 通过移除 setup-automation.sh 脚本,定时提醒的配置方式由自动脚本改为手动编辑 crontab。 - 文档更新:SKILL.md 中相关说明同步修改,详细说明如何用 crontab 启用每天 9:30 钉钉提醒。 - 删除无关的配置文件(.clawhub.json, .clawhub/origin.json),进一步精简项目结构。 - 其余核心执行逻辑、用法和接口保持不变。
v1.1.0
规则存储优化 - 按日期合并文件,每天 1 个文件 定时提醒功能 - 钉钉自动推送待确认规则(默认 9:30) 规则状态管理 - 待审/已确认状态正确显示 完整日志记录 - 同意/忽略操作都记录日志 通用化配置 - 用户自行配置钉钉账号和提醒时间 配置统一 - 所有配置集中在 config/config.yaml 文档完善 - 添加历史文档整合说明
v1.0.1
Initial release. - Implements core self-improvement processes: error prevention before each reply, self-checking and draft rule generation after each reply, and automatic memory cleanup every 10 rounds or weekly. - Provides strong safety principles: human approval required for any persistent memory change, AI can only propose rules, and never modifies or deletes confirmed rules automatically. - Includes detailed memory structure: separate areas for dialogue count, proposal drafts, logs, and confirmed rules. - Enforces unified rule draft format and credibility requirements. - Adds example scripts and usage scenarios for integration. - Supplies LICENSE and SECURITY guidelines.
v1.0.0
developer-self-improve-core 1.0.0 - Introduces a self-improving core skill for developers, focused on safe, rules-based automation with user approval. - Adds automatic pre-response error prevention, post-response self-check with rule proposal, and periodic memory cleanup (every 10 rounds or weekly). - Enforces strict safety: AI can only propose—never auto-save or modify long-term memory; user review is always required. - Provides standardized formats for rule proposals, strong traceability, rollback, and memory layering (short-term, proposal, long-term, cleanup, logs). - Includes sample workflows, command-line scripts for each core operation, and detailed documentation links.
Metadata
Slug developer-self-improve-core
Version 1.1.9
License MIT-0
All-time Installs 1
Active Installs 1
Total Versions 12
Frequently Asked Questions

What is developer-self-improve-core?

开发者自改进核心技能 - 自动错误防重、自检、规则生成、记忆清洗、定时提醒 核心功能: - 每轮回答前:自动错误防重 - 每轮回答后:自动自检 + 生成规则草案 - 累计 10 轮对话/每周:自动记忆清洗扫描 - 自动提醒:每天 9:30 钉钉推送待确认规则 核心原则: - AI 只提议,人类终审 - 绝不自动写... It is an AI Agent Skill for Claude Code / OpenClaw, with 164 downloads so far.

How do I install developer-self-improve-core?

Run "/install developer-self-improve-core" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is developer-self-improve-core free?

Yes, developer-self-improve-core is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does developer-self-improve-core support?

developer-self-improve-core is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created developer-self-improve-core?

It is built and maintained by Joe.Lee (@joelee09); the current version is v1.1.9.

💬 Comments