A meta-skill that makes AI agent skills evolve — create, reflect, evaluate, publish, search, install, fork, merge, review, and uninstall skills autonomously.
Not another skill marketplace. This is the engine that makes skills on any marketplace learn and improve.
让 AI 技能自己会长的系统——创建、反思、评测、发布、搜索、安装、Fork、合并、评审、卸载，全自动。
不是又一个技能市场，而是让任何市场上的技能都能学习和进化的引擎。
⚠️ This is NOT a human-facing app. There's no UI, no sign-up, no login.This is a skill for Claude Code (an AI coding agent). You give it a GitHub link, the AI clones it, and it works. The "user" is the AI, not you.
⚠️ 这不是给人操作的应用。没有界面、没有注册、没有登录。这是给 Claude Code（AI 编程 agent）用的技能。你把这个 GitHub 链接丢给 Claude Code，AI 自己会 clone 安装。使用者是 AI，不是你。
How to use | 怎么用： Open Claude Code → paste the GitHub URL of this repo → AI handles the rest.
怎么用： 打开 Claude Code → 把这个仓库的 GitHub 链接丢给它 → AI 自己搞定。
- Claude Code (or any AI coding agent that supports
.claude/skills/) - Python 3.9+
No pip packages required. All scripts use Python standard library only.
无需 pip install 任何依赖。所有脚本仅使用 Python 标准库。
cd your-project
mkdir -p .claude/skills
git clone https://github.com/hao-cyber/skill-evolution.git .claude/skills/skill-devYour project should now look like | 项目结构：
your-project/
├── .claude/
│ └── skills/
│ └── skill-dev/
│ ├── SKILL.md
│ ├── scripts/
│ └── references/
└── ...
Done. Claude Code will automatically pick up skill-dev.
搞定。 Claude Code 会自动发现 skill-dev，立即可用。
No .env, no accounts, no setup. The public registry is built in.
不需要 .env，不需要注册账号，不需要配置服务器。公共市场已内置。
# Search | 搜索
python3 .claude/skills/skill-dev/scripts/search.py --query "web scraper"
# Install (dependencies auto-installed) | 安装（自动安装依赖）
python3 .claude/skills/skill-dev/scripts/install.py --name web-read
# Publish (preview first, then --yes to confirm) | 发布（先预览，再确认）
# Publisher identity is auto-generated on first publish.
# 发布者身份在首次发布时自动生成。
python3 .claude/skills/skill-dev/scripts/publish.py --skill-name my-skill
python3 .claude/skills/skill-dev/scripts/publish.py --skill-name my-skill --yesTo run a separate registry for your team | 为团队搭建私有市场：
a. Create a free Supabase project | 创建免费 Supabase 项目
b. Run setup.sql in the Supabase SQL Editor | 在 SQL 编辑器中运行 setup.sql
c. Set SUPABASE_URL and SUPABASE_ANON_KEY env vars | 设置环境变量覆盖默认值
Even without any network, skill-dev gives your agent:
即使没有网络，skill-dev 也能让你的 AI：
- Skill creation — agent discovers a capability gap and builds a new skill | 创建技能 — AI 发现能力缺口，自动构建新技能
- Skill reflection — after failure, agent analyzes root cause and fixes the skill | 反思修复 — 执行失败后，AI 分析根因并自动修复
- Maturity assessment — after reflection cycles stabilize, agent proactively suggests publishing | 成熟度判断 — 反思周期稳定后，AI 主动建议发布
- Skill structure — enforces standard format (YAML frontmatter, progressive loading) | 结构规范 — 强制标准格式（YAML frontmatter、渐进式加载）
The public registry adds publish/search/install on top. The core value works 100% offline.
公共市场在此之上提供发布/搜索/安装。核心价值 100% 离线可用。
web-scraper/
├── SKILL.md # When to use, how to do it right | 何时触发、怎么做对
├── scripts/ # Deterministic code (optional) | 确定性代码（可选）
├── references/ # Deep docs, loaded on demand (optional) | 深度文档，按需加载
└── assets/ # Templates, materials (optional) | 模板、素材
Progressive loading keeps context costs flat | 渐进式加载让上下文开销趋近于零：
- Metadata (name + description) — always in context (~100 words) | 元数据 — 始终在上下文
- SKILL.md body — loaded when skill triggers | SKILL.md 正文 — 触发时加载
- references/ — loaded only for specific scenarios | 参考文档 — 仅特定场景加载
Installing 50 skills costs the same context as installing 1 — until one is actually needed.
装 50 个技能和装 1 个，上下文开销一样——直到某个技能真正被触发。
No semver. Skills fork into named variants | 不用版本号，技能 Fork 为命名变体：
web-scraper (base) <- Original | 原版
├── web-scraper@alice <- Added proxy rotation | 加了代理轮换
├── web-scraper@bob <- Parallel execution | 加了并发执行
└── web-scraper@merged <- Agent-merged best of both | AI 合并最优版
When you publish a skill that already exists under a different author, it automatically forks as a new variant. Agents choose the best variant based on task context (audited > description match > installs > review score).
发布已存在的同名技能时，自动 Fork 为新变体。AI 根据任务上下文自动选最优版本（审计通过 > 描述匹配 > 安装量 > 评分）。
Publishes a local skill to the registry. Defaults to preview mode — requires --yes to upload.
发布本地技能到市场。默认预览模式 — 需要 --yes 才会真正上传。
--skill-name NAME Skill directory name under .claude/skills/
--variant VARIANT Variant name (default: base)
--author AUTHOR Author identifier (default: git config user.name)
--yes Actually publish (without this, only preview is shown)
Searches the registry | 搜索市场。
--query KEYWORDS Full-text search | 全文搜索
--tag TAG Filter by tag | 按标签过滤
--sort ORDER Sort by: installs (default), updated, name
--limit N Max results (default: 10)
--offset N Skip first N results for pagination
--detail NAME Show all variants for a specific skill | 查看某技能所有变体
--list-all List everything | 列出全部
--include-unaudited Include skills that haven't passed security audit | 包含未审计技能
Downloads and installs a skill from the registry | 从市场安装技能。
--name NAME Skill to install
--variant VARIANT Variant to install (default: base)
--force Overwrite existing skill directory | 覆盖已安装的同名技能
--no-deps Skip automatic dependency installation | 跳过自动依赖安装
Removes a locally installed skill | 卸载本地技能。
--name NAME Skill to uninstall
--yes Skip confirmation | 跳过确认直接删除
Scaffold for merging two skill variants. Agent handles the semantic merge; script handles the plumbing.
变体合并脚手架。AI 负责语义合并，脚本负责管道工作。
merge.py prepare --name NAME --variants a,b [--workspace DIR]
merge.py diff --dir-a PATH --dir-b PATH
merge.py publish --workspace DIR --name NAME [--variant merged] [--yes]
Submit and view skill reviews. Publisher identity is auto-managed.
提交和查看技能评价。发布者身份自动管理。
review.py submit --skill-name NAME --score 1-5 [--review "text"] [--context "context"]
review.py list --skill-name NAME [--variant base] [--limit N]
review.py stats --skill-name NAME
Security audit scanner (admin-only) | 安全审计扫描器（仅管理员）。
--name NAME Audit specific skill (default: all)
--dry-run Show findings without updating DB
--verbose Show detailed findings per skill
- Admin (service_role key): audit skills, reset publisher keys, full DB access | 管理员：审计技能、重置密钥、完整数据库访问
- Users (anon key + publisher key): search, install, publish own skills, submit reviews | 用户：搜索、安装、发布自己的技能、提交评价
- All writes go through security-definer RPCs — anon key cannot write directly | 所有写操作通过安全 RPC — 匿名密钥无法直接写入
- Publisher identity prevents author impersonation | 发布者身份防止冒充
- Updated skills have
audited_atcleared — must be re-audited | 更新后的技能需重新审计
Agent <-> skill-dev (SKILL.md + scripts/)
|
Supabase (PostgreSQL)
|
skills table + full-text search
All complex decisions (which variant to pick, how to merge, quality assessment) are made by the agent. Infrastructure just stores and queries.
所有复杂判断（选版本、合并、质量评估）由 AI 做。基础设施只管存和查。
- Works without registry | 离线可用：Local skill creation and reflection need zero infrastructure | 本地创建和反思不需要任何基础设施
- For agents, not humans | 为 AI 设计：Pure API, no web UI needed | 纯 API，不需要 Web 界面
- Skills evolve | 技能会进化：Fork, personalize, merge — agents choose the best version | Fork、个性化、合并 — AI 自动选最优
- Zero infrastructure | 零基础设施：Public registry built-in, works out of the box | 公共市场内置，开箱即用
MIT