10xAI Wiki

12分钟阅读

Claude Skills & MCP Protocol

Two Parallel Extension Systems

The AI agent ecosystem has converged on two complementary approaches for extending AI model capabilities: Skills (prompt-based capability modules) and MCP (Model Context Protocol — standardized tool integration). Understanding when to use each is key to building effective AI workflows.

Claude Skills

What Is a Skill?

A Skill is a reusable capability module: a prompt template + optional scripts that extends an AI model’s ability to perform specific tasks. In Claude Code and similar tools, skills are triggered by:

  • Slash commands (e.g., /commit, /review)
  • Natural language auto-matching (the model recognizes the intent and applies the skill)

Skills do not invent new tools — they compose existing tools in smart ways. They are essentially “canned expertise.”

Notable Skills (Real-World Examples)

ai-translate — Embedded Translation Skill

stormzhang’s open-source translation skill embedded in AI coding tools. Two characters (/t) trigger translation, no API key required. Key design principles:

  • Context Fork mode ignores project context, minimizing token cost
  • Zero-config: reuses the AI model already powering the coding tool (Claude or GPT)
  • Five tool adapters: Claude Code, Codex, Cursor, Windsurf, OpenCode
  • Voice variant: /ts translates and reads aloud
  • Context-aware: when translated word is a coding tool command, provides usage help

See: ../sources/2026-05-05-我开源了一个 AI 轻量翻译工具

ai-translate exemplifies the “Prompts-as-Products” philosophy — two carefully designed prompt files constitute the entire product. Competitive moat = prompt quality + “convenient enough.”

Top Skills (from 50,000+ user collection)

SkillPurposePopularity
Skill MentorLearning prompt engineeringVery High
Claude Code Best PracticesUsing Claude Code effectivelyVery High
Karpathy’s AutoresearchIterative prompt optimizationHigh
Code ReviewStructured code reviewHigh
Meeting SummaryAuto-generate meeting notesHigh

Skill Marketplace

ClawHub (for OpenClaw) and similar platforms allow:

  • Publishing skills to community
  • Browsing by category/use case
  • One-click installation
  • Version management

YAML skill configuration example:

name: code-review
description: Structured code review with security focus
trigger: /review
instructions: |
  Review the provided code for:
  1. Security vulnerabilities
  2. Performance issues
  3. Code style consistency
  4. Test coverage
tools: [Read, Grep, Glob, Bash]

The “Skill Designer Engineer” — New Role?

One article predicts a new job title emerging: Skill Designer Engineer

  • Crafting effective prompts that work across many users and contexts
  • Understanding which tool combinations solve specific task types
  • Testing and iterating on skill effectiveness
  • Similar to prompt engineering but at scale

Superpowers — 生产级 Workflow Skill

Superpowers 是 Claude Code 的生产级工作流 Skill（../sources/2026-05-02-Claude-Code-MiniMax-Superpowers-生产系统交付）。三步工作流：

  1. /brainstorming：苏格拉底式引导讨论，主动收集必要信息，补全纠正需求
  2. /writing-plans：编写结构化执行计划文件，待用户确认
  3. /subagent-driven-development：启动子 Agent 按步骤执行计划，自动 git commit 保持可追溯性

配套工具：

  • Planning-with-files：3 个文件约束长任务失控问题，适合跨天任务
  • Claude-hud：实时展示上下文用量，超 70% 建议 compact 或新会话

安装：

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

Skill 工程方法论（2026-06 更新）

Source: 2026-06-14-Skill-Engineering-Methodology

Skill = Context Engineering（Anthropic 核心观点）

Anthropic 内部 Skill 方法论的关键认知转变：

  1. Skill 不是一个 markdown 文件，而是一个文件夹：用文件夹形式表达 Context Engineering 理念
  2. 渐进式暴露：SKILL.md 是导航页，references/、scripts/、examples/、assets/ 按需加载
  3. 三层加载 Token 经济学：
    • Level 1（常驻）：name + description，~50-150 Token/Skill
    • Level 2（触发加载）：SKILL.md 正文，~2,000-5,000 Token
    • Level 3（按需）：脚本/参考文档，按实际大小
  4. Gotchas > Instructions：Skill 真正要写的是常踩的坑（“这个表不能按 created_at 排序”），不是模型已知的常识

Instructions vs Scripts 分界

层面InstructionsScripts
作用经验和判断能力和执行
示例”Stripe 返回 200 不代表成功”check_payment_events()

Description = 路由规则

Description 最重要的不是”能干什么”，而是”什么情况下应该加载它”。检验方法：删掉 Skill，只保留 Description，模型能否判断何时触发。

常见反模式（腾讯技术工程总结）

反模式症状解法
大杂烩一个 Skill 干三件事拆分为子 Skill
Description 黑话全是内部术语通用语言+技术关键词
没有示例纯文字描述Few-Shot + Before/After
没有验证点做完才检查关键步骤加检查点
当 Wiki 写背景 300 行正文 50 行背景放 references/

Skill Creator 工程化评估

Anthropic 官方 Skill Creator 新增评估能力：触发评估（Precision/Recall ≥ 85%）+ 效果评估（通过率 ≥ 80%，相对提升 ≥ 30%）。

鲁班 Skill：Skill 自我升级

用 Claude Fable 5 打磨升级已有 Skill。横纵对比法：分析 GitHub 同类项目 Star 历史和 Release 记录，找出差距功能，自动迭代。

MCP vs HTTP 选择决策树

需要调用外部服务
  → 已有 MCP Server？ → 优先 MCP
  → 需多平台复用？ → 封装 MCP Server  
  → 简单一次性调用？ → 脚本直接 HTTP

核心原则：MCP 管连接，Skill 管流程，HTTP 脚本兜底

Enterprise Pipeline Skills: NVIDIA AI-Q

NVIDIA AI-Q expands the skill concept from “prompt + helper script” to “thin adapter over a full backend pipeline.” The AI-Q skill gives Claude Code, Codex, OpenCode, or LangChain Deep Agents a single deep-research capability backed by a running AI-Q server.

Key design distinction:

  • Skill layer: tells the harness when/how to delegate a research task and uses a helper script to submit and poll jobs.
  • MCP layer: connects enterprise data sources to AI-Q with auth patterns: unauthenticated mcp_client, service-account mcp_client + mcp_service_account, or signed-in user bearer-token forwarding.
  • Backend layer: AI-Q handles intent classification, clarification, shallow/deep research, evaluation, citations, and OpenTelemetry traces.

This is a useful layered pattern: Skills expose high-level reusable capabilities to the agent; MCP supplies data/tool connectivity; the specialized backend owns reliability, security, and evaluation.

MCP — Model Context Protocol

What Is MCP?

MCP (Model Context Protocol) is an open standard invented by Anthropic for connecting AI models to external tools and data sources. Think of it as “USB-C for AI” — a universal connector that works across different AI providers and tools.

Core concept: Instead of each AI model implementing custom integrations for every tool (n×m problem), MCP defines a standard interface. Tool developers implement the standard once; AI models can use any MCP-compatible tool.

Architecture

┌─────────────┐     MCP Server      ┌──────────────┐
│ AI Model    │ ←────────────────→ │ Tool (GitHub) │
│ (Claude)    │   (Standard Protocol)│ (Database)   │
└─────────────┘                     └──────────────┘
              ↑─────────────────↓
        MCP Server (Brave Search)

Key MCP Servers (Community-Owned)

ServerCapabilities
GitHubRead/write repos, issues, PRs
Google DriveRead/write docs, sheets
Slack/DiscordSend/receive messages
PostgreSQLQuery databases
Brave SearchWeb search
FilesystemRead/write local files
PuppeteerBrowser automation

MCP in Claude Code

Claude Code supports both MCP client (connecting to external MCP servers) and MCP server (exposing Claude Code tools to other applications).

# Example: Add GitHub MCP to Claude Code
claude mcp add github -- npx @anthropic/claude-mcp-server-github

Skills vs MCP — When to Use Each

DimensionSkillsMCP
What it doesPrompt template + tool compositionTool integration standard
FlexibilityHigh — can encode complex logicFixed — follows protocol spec
Context impactAdds to context windowCan pollute context if misused
MaintenancePrompt versioningServer maintenance
Best forComplex reasoning patterns, specialized expertiseExternal data access, API calls
ComposableCan call MCP tools inside a skillCan be triggered by skills

Peter Steinberger’s View (OpenClaw Creator)

“MCP 污染上下文，不可组合；Skills + CLI 才是正道”

His argument:

  • MCP pulls entire data blobs into context
  • Skills + CLI: model calls Unix commands, filters with jq, gets only what it needs
  • No context pollution
  • More efficient and composable

The Emerging Consensus

Use MCP for: Data retrieval, API integrations, external services Use Skills for: Complex workflows, domain expertise, multi-step reasoning patterns Use both together: Skill invokes MCP tool calls

Obsidian Skills — Knowledge Management Extensions

Obsidian has developed its own Skill system for AI-powered knowledge management:

  • Defuddle: Extract content from web pages (Obsidian CEO’s own skill)
  • Memory: Persistent context across conversations
  • Search: Semantic search across the vault

These represent a new paradigm: Skills that make AI understand your knowledge base.

Sources: Claude Agent Skills Deep Dive, Skill Designer Engineer, MCP vs Skills vs PTC, Obsidian Skills, Top 10 Skills

关系图谱

反向链接