Hooks, skills, and MCP as a build system

How do you extend Claude Code with hooks, skills, and MCP servers?

Claude Code gives you three extension points and they do different jobs: hooks fire on the agent's lifecycle so you can gate what it is about to do, skills carry the procedure you want followed instead of whatever it remembers, and an MCP server hands it tools and resources it could not otherwise reach. Guardian Kane uses all three, a Stop hook that grew into a twelve phase loop, a skill holding the procedure, and a kane-cli underneath, with a four tab dashboard showing the code graph, memory graph, PRD graph, and live activity. I have written up what the second build taught me about how much of it is wiring.