本文由 AI 分析生成
建立時間： 2026-03-28 來源： https://gist.github.com/coodoo/964ba1d813049e94b6d7680dcc4edfd3
Summary
Coodoo’s GitHub Gist (Traditional Chinese) clarifying the distinction between MCP (Model Context Protocol) tools and Claude Code Skills, and why Skills were created. Core argument: MCP tools load their full spec into context at startup, have high per-call overhead, and can’t easily chain multi-tool workflows. Skills solve these by lazily loading only needed specs and enabling multi-step code execution in a single call.
Coodoo 的 GitHub Gist（正體中文）澄清 MCP 工具與 Claude Code Skills 的區別，以及 Skills 被創建的原因。核心論點：MCP 工具在啟動時將完整規範加載到上下文中，每次調用開銷高，且無法輕易鏈接多工具工作流程。Skills 通過延遲加載所需規範並在單次調用中啟用多步驟代碼執行來解決這些問題。
Key Points
- Function calling = the fundamental mechanism for all tool use (MCP, skills, bash, read, write — all are function calls)
- MCP problems: (1) all MCP specs loaded at startup → burns context; (2) each call returns intermediate data to Claude → more context use; (3) can’t flexibly combine multiple MCPs in one operation
- Skills solution: (1) lazy loading — only load the skill spec when invoked; (2) code-based composition — skills can invoke multiple MCPs programmatically in a single turn; (3) enable workflows like “query context7 + deepwiki + github simultaneously”
- Practical example: querying 3 MCPs (context7, deepwiki, github) for React docs: MCP method = 4 context-consuming round trips; Skill method = 1 call that internally runs all 3 and returns combined result
- Skills leverage increasingly capable LLM coding abilities to implement flexible composition that MCP’s sequential model can’t match
Insights
This is an important explanation for why Claude Code uses Skills rather than pure MCP orchestration for complex workflows. The key insight is that context is a finite resource: loading every MCP’s full specification at session start, and returning intermediate results for each sub-step, compounds context waste. Skills invert this: the spec loads only when called, and the skill itself handles intermediate state in code rather than in the context window. This analysis implicitly predicts that as models become more capable coders, the “skill as code” pattern will increasingly dominate over declarative MCP-style tool calling for complex workflows.
Connections
Raw Excerpt
Skills 的出現就是為了解決上述三個主要缺點，並結合 LLM 日漸強大的寫扣能力帶來更彈性的應用可能性。單次操作模式且每次需返還中間暫存資料佔用大量 context。