skila

Self-improving skill manager for Claude Code

Automatically distill coding sessions into reusable, versioned skills — and continuously refine them as you work.

English | 中文

English

Install

# npm (global CLI)
npm i -g @luyao618/skila

# Claude Code plugin
/plugin marketplace add luyao618/skila
/plugin install skila@skila

Quick start

# run inside Claude Code — distill current session into a reusable skill
/skila

# run in terminal — launch web control panel
skila serve

How it works

/skila turns a Claude Code session into a complete skill directory — automatically.

Here's /skila in action — distilling a debugging session into a new skill and updating an existing one:

The pipeline has three stages:

1. Rule-based extraction — scans tool traces for reusable artifacts. Repeated/complex Bash → scripts/, read docs → references/, written templates → assets/.
2. LLM Judge — decides NEW or UPDATE against existing inventory. Filters and supplements extracted files. Falls back to heuristic matching when no API key is available.
3. Write + validate — outputs a spec-compliant skill directory with hallucination guards and path safety checks.

Web control panel

skila serve launches a three-pane Obsidian-style dashboard:

• Skill editor — CodeMirror 6 markdown editor with raw / preview / translate views
• LLM translation — translate skills to any language via streaming SSE (Claude API)
• Version control — browse version history, rollback to any previous version
• Lifecycle management — filter and transition skills across statuses: draft → staging → published → archived / disabled
• Feedback inspector — success rate, usage count, invocation history

Inspired by

• hermes-agent — autonomous agent with self-managed skill lifecycle
• memex — Zettelkasten-based persistent knowledge system

License

MIT © yao 2026

中文

安装

# npm 全局安装
npm i -g @luyao618/skila

# Claude Code 插件
/plugin marketplace add luyao618/skila
/plugin install skila@skila

快速开始

# 在 Claude Code 中运行 — 将当前会话提炼为可复用的 skill
/skila

# 在终端中运行 — 启动 Web 控制面板
skila serve

工作原理

/skila 将 Claude Code 会话自动转化为完整的 skill 目录。

下图展示了 /skila 的实际运行效果 — 从调试会话中提炼新 skill 并更新已有 skill：

流程分三个阶段：

1. 规则提取 — 扫描工具调用记录，提取可复用产物。重复/复杂的 Bash 命令 → scripts/，读取的文档 → references/，写入的模板 → assets/。
2. LLM 审核 — 对比现有 skill 库，判断新建还是更新。过滤和补充提取的文件。无 API Key 时自动降级为启发式匹配。
3. 写入 + 验证 — 输出符合规范的 skill 目录，内置幻觉防护和路径安全检查。

Web 控制面板

skila serve 启动三栏 Obsidian 风格的管理面板：

• Skill 编辑器 — CodeMirror 6 Markdown 编辑器，支持原文 / 预览 / 翻译视图
• LLM 翻译 — 通过 Claude API 流式翻译 skill 到任意语言
• 版本控制 — 浏览版本历史，回滚到任意历史版本
• 生命周期管理 — 按状态筛选和流转 skill：draft → staging → published → archived / disabled
• 反馈面板 — 成功率、使用次数、调用历史

灵感来源

• hermes-agent — 具有自管理 skill 生命周期的自主 agent
• memex — 基于 Zettelkasten 的持久化知识系统

许可证

MIT © yao 2026