create-expert-skill

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Expert Skill Creation

专家级Skill创建

Transform expert knowledge into production-ready skills that combine domain expertise with system-specific understanding.

将专家知识转化为兼具领域专业能力与系统专属认知的、可用于生产环境的Skill。

Why Skills Fail in Production

为何Skill在生产环境中失效

AI assistants fail not because they lack intelligence, but because they lack:

Domain Expertise — Industry-specific rules, edge cases, unwritten conventions
Ontology Understanding — How YOUR systems, data structures, and workflows actually work

Both are required. Domain knowledge without system context produces generic output. System knowledge without domain expertise produces structurally correct but semantically wrong results.

AI助手并非因缺乏智能而失效，而是因为缺少：

领域专业知识 —— 行业特定规则、边缘案例、不成文惯例
本体认知 —— 您的系统、数据结构与工作流的实际运作方式

二者缺一不可。脱离系统语境的领域知识会产生通用化输出；缺乏领域专业知识的系统认知则会生成结构正确但语义错误的结果。

Workflow

工作流

Assess → Discover (Expertise + Ontology) → Design → Create → Refine → Ship

评估 → 挖掘（专业知识 + 本体）→ 设计 → 创建 → 优化 → 发布

Quick Assessment

快速评估

Create a skill when:

Used 3+ times (or will be)
Follows consistent procedure
Saves >300 tokens per use
Requires specialized knowledge not in Claude's training
Must produce trusted output (not "close enough")

Don't create for: one-time tasks, basic knowledge Claude already has, rapidly changing content.

适合创建Skill的场景：

已使用3次及以上（或将会被使用这么多次）
遵循固定流程
每次使用可节省300个以上tokens
需要Claude训练数据中未涵盖的专业知识
必须生成可信任的输出（而非“近似即可”）

**不适合创建的场景：**一次性任务、Claude已掌握的基础常识、内容快速变化的场景。

Discovery: Two Streams

挖掘：两大方向

Stream 1: Domain Expertise

方向1：领域专业知识

Deep knowledge that transcends any specific company:

Industry standards and their versions
Professional conventions and best practices
Edge cases only practitioners know
Validation rules from specifications

Example (LEDES validation): LEDES 98B vs XML 2.0 formats, UTBMS code taxonomy, date format requirements, required vs optional fields.

超越特定企业的深度知识：

行业标准及其版本
专业惯例与最佳实践
只有从业者才了解的边缘案例
规范中的验证规则

示例（LEDES验证）： LEDES 98B与XML 2.0格式差异、UTBMS代码分类体系、日期格式要求、必填与可选字段。

Stream 2: Ontology Understanding

方向2：本体认知

How the skill maps to specific systems and organizations:

Company-specific policies and constraints
Data structures and identifiers unique to the system
Cross-references between entities (timekeepers → IDs → rates)
Workflow states and transitions

Example (LEDES validation): Firm-specific timekeeper codes, matter numbering conventions, approved billing rates, outside counsel guideline requirements.

Skill如何适配特定系统与组织：

企业专属政策与约束
系统特有的数据结构与标识符
实体间的交叉引用（计时员 → ID → 费率）
工作流状态与转换规则

示例（LEDES验证）： 律所专属计时员代码、事务编号惯例、核准的计费费率、外部法律顾问准则要求。

Discovery Questions

挖掘问题清单

When starting, I'll ask about:

Domain & Purpose — What problem? What industry standards apply?
Ontology Requirements — What system-specific structures must the skill understand?
Content Source — Conversation, docs, specifications, or files to distill from?
Automation Potential — What can be deterministic (scripts)? What needs interpretation (LLM)?
Complexity Level — Simple (SKILL.md only), Enhanced (+scripts), or Full (+resources)?

启动时，我会询问以下问题：

领域与目标 —— 要解决什么问题？适用哪些行业标准？
本体要求 —— Skill必须理解哪些系统专属结构？
内容来源 —— 从对话、文档、规范还是文件中提炼？
自动化潜力 —— 哪些工作可通过脚本实现确定性执行？哪些需要LLM进行解读？
复杂程度 —— 简单版（仅SKILL.md）、增强版（+脚本）还是完整版（+资源）？

Skill Architecture

Skill架构

skill-name/
├── SKILL.md              # Layer 1: Core (300-500 tokens)
├── scripts/              # Layer 0: Automation (0 tokens to run)
│   └── validate.py
└── resources/            # Layer 2: Details (loaded selectively)
    └── ADVANCED.md

Layer 0 (Scripts): Free execution, structured JSON output Layer 1 (SKILL.md): Loaded when triggered - keep lean Layer 2 (Resources): Fetched only when specific section needed

skill-name/
├── SKILL.md              # 第一层：核心内容（300-500 tokens）
├── scripts/              # 第零层：自动化（运行时消耗0 tokens）
│   └── validate.py
└── resources/            # 第二层：细节内容（按需加载）
    └── ADVANCED.md

第零层（脚本）：执行无消耗，输出结构化JSON 第一层（SKILL.md）：触发时加载 - 保持精简 第二层（资源）：仅在需要特定章节时调用

Token Optimization

Token优化策略

Technique	Instead of	Do this	Savings
Scripts	500 tokens explaining validation	`python scripts/validate.py`	~450 tokens
Reference	Inline schema (200 tokens)	Link to `resources/schema.json`	~185 tokens
Layer 2	Everything in SKILL.md	Link to `resources/ADVANCED.md`	~750 tokens

技巧	替代方案	优化方式	节省量
脚本	用500 tokens解释验证逻辑	`python scripts/validate.py`	~450 tokens
引用	内联模式（200 tokens）	链接至 `resources/schema.json`	~185 tokens
分层加载	所有内容放入SKILL.md	链接至 `resources/ADVANCED.md`	~750 tokens

Description Formula

描述公式

<Action> <Object> for <Purpose>. Use when <Trigger>.

Example: "Validate billing data for system migration. Use before importing invoices."

<动作> <对象> 以实现 <目标>。适用于 <触发场景>。

示例：“验证计费数据以支持系统迁移。适用于导入发票前的校验环节。”

Shipping

发布流程

When content is finalized:

bash

python scripts/package_skill.py skill-name 1.0

Creates

skill-name-v1.0.zip

with:

DIRECTORY_STRUCTURE.txt (auto-generated)
README.md with deployment instructions
All skill files properly organized

内容定稿后执行：