本文由 AI 分析生成
建立時間： 2026-01-13 來源： https://x.com/eyad_khrais/status/2010810802023141688
Summary
Part two of a Claude Code tutorial covering the three advanced systems most users underuse: Skills (markdown files with name/description frontmatter that Claude auto-applies via progressive disclosure), Subagents (isolated Claude instances with their own context, system prompt, and tool permissions that communicate only via delegation-and-summary), and MCP connectors (a standardized interface to external tools like GitHub, Slack, and databases). It argues these compound into a system that multiplies engineering work, and notes Claude Code delivers a consistent 200K context window where other tools silently truncate.
Claude Code 教學第二部，講解多數人未善用的三大進階系統：Skills（帶 name/description frontmatter 的 markdown，透過漸進式揭露自動觸發）、Subagents（擁有獨立 context、系統提示與工具權限的子實例，僅透過「委派並回傳摘要」溝通）、以及 MCP 連接器（連到 GitHub、Slack、資料庫等外部工具的標準化介面）。作者主張三者疊加能成倍放大工程產出，並指出 Claude Code 提供穩定的 200K context，而其他工具會悄悄截斷。
Key Points
- Skills live in
~/.claude/skills/<name>/SKILL.md; the description field is what Claude uses to decide when to apply them. - Progressive disclosure: only name + description (~100 tokens each) preload; full instructions load only when relevant, so dozens of skills don’t bloat context.
- Built-in subagents are Explore (read-only search), Plan (research for plan mode), and General-purpose (multi-step tasks); custom ones go in
~/.claude/agents/. - Subagents have isolated 200K context and communicate via delegation + summary; they cannot spawn other subagents (no infinite nesting).
- Context degradation begins around 45% of the window — subagents offload work to fresh context to keep the main conversation clean.
- MCP add command:
claude mcp add --transport http <name> <url>; collapses multi-tool workflows into one continuous session.
Insights
The framing that “the documentation doesn’t tell you how the pieces fit together” is the real value: the article treats skills, subagents, and MCP as one composable architecture (encode conventions → isolate subtasks → eliminate context switching) rather than three separate features. The explicit warning that the delegation summary — not shared state — is the communication channel explains why subagent prompts must specify output format, a subtlety that directly informs how bot_vault’s own skills should be authored. Note the third-party MCP trust caveat.
Connections
Raw Excerpt
A skill that knows your codebase patterns + a subagent that handles testing + MCP connections to your issue tracker = a system that is unmatched.