Six total. Here's the full stack and how it composes.
𝗧𝗵𝗲 𝘁𝘄𝗼 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝘁𝗮𝗹𝗸𝘀 𝗮𝗯𝗼𝘂𝘁
𝗦𝗸𝗶𝗹𝗹𝘀 = 𝗞𝗡𝗢𝗪𝗟𝗘𝗗𝗚𝗘 (𝘄𝗵𝗮𝘁 𝘁𝗼 𝗱𝗼)
→ Reusable instruction modules in .claude/skills/ → Each SKILL. md loaded on-demand via progressive disclosure → Like training manuals the agent reaches for when relevant
𝗠𝗖𝗣 = 𝗔𝗕𝗜𝗟𝗜𝗧𝗬 (𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝘀 𝘁𝗼 𝘁𝗵𝗲 𝗲𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝘄𝗼𝗿𝗹𝗱)
→ GitHub, Slack, AWS, Notion, databases — anything with an API → 10,000+ servers, now under Linux Foundation governance → 2026: MCP Apps bring interactive UI into the loop
Simon Willison thinks Skills might be a bigger deal than MCP. Maybe. But neither matters without the four below.
𝗧𝗵𝗲 𝗳𝗼𝘂𝗿 𝘁𝗵𝗮𝘁 𝗺𝗮𝗸𝗲 𝗶𝘁 𝘄𝗼𝗿𝗸
𝗦𝘂𝗯𝗮𝗴𝗲𝗻𝘁𝘀 = 𝗗𝗘𝗟𝗘𝗚𝗔𝗧𝗜𝗢𝗡
→ Isolated context, own model, own permissions → Code Reviewer, Researcher, Deployer — each a specialist → Why your main context window doesn't drown in noise
𝗛𝗼𝗼𝗸𝘀 = 𝗔𝗨𝗧𝗢𝗠𝗔𝗧𝗜𝗢𝗡
→ PRE-TOOL fires before tool execution, POST-TOOL after → ON-EDIT fires when files change → Deterministic — NOT controlled by the LLM → Where you enforce things the model can't be trusted to remember
𝗖𝗟𝗔𝗨𝗗𝗘.𝗺𝗱 = 𝗔𝗟𝗪𝗔𝗬𝗦-𝗢𝗡 𝗖𝗢𝗡𝗧𝗘𝗫𝗧
→ Project conventions, architecture, company info → Loaded automatically into every session → The floor every other primitive builds on
𝗣𝗹𝘂𝗴𝗶𝗻𝘀 = 𝗣𝗔𝗖𝗞𝗔𝗚𝗜𝗡𝗚
→ Bundle Skills + Hooks + Subagents + MCP into one installable unit → What turns a personal setup into something a team can adopt
𝗛𝗼𝘄 𝗶𝘁 𝗰𝗼𝗺𝗽𝗼𝘀𝗲𝘀
A "competitive analysis" request hits Claude Code:
- CLAUDE. md loads project + company context
- The competitive-analysis Skill activates the framework
- MCP fires → searches Drive for past briefs
- Subagent spawns → market researcher gathers data
- Subagent spawns → technical analyst reviews competitor repos
- Hook triggers → auto-formats output, runs linter
One request. Six primitives. Skills and MCP are what you see in the trace. The other four are what made the trace possible.
Skills tell the agent what good work looks like. MCP gives it reach. Subagents are what let it scale without losing context. Hooks are what keep it honest. CLAUDE. md is what stops it forgetting where it is. Plugins are what let your whole team run the same playbook.
Pull any of the four and the visible two collapse.
So back to Simon's claim — Skills bigger than MCP? Or is the real argument that neither of the visible two matters as much as the invisible four nobody's talking about?
#jamesCodeLab #MultimodalAI #RAG #fblifestyle #AIEngineering #LLMSystems #GeminiAI #PerplexityAI #AgenticAI
