If your Claude Code setup feels messy, it's usually because one of these is doing a job that belongs to another.
They're five different tools for five different problems:
- CLAUDE.md: context that's always on. Who you are, your rules.
- Skills: instructions that load only when a task needs them.
- Subagents: isolate heavy work so it doesn't crowd the main thread.
- Hooks: run something automatically at a set point in the loop.
- MCP: connect Claude to outside tools and data.
Pick the wrong one and you either bloat every prompt or bolt on complexity you don't need. Reach for the simplest piece that solves the actual problem, and add the others only when the task asks for it.
New video walks through each one, when to use it, and when not to.
Which of these have you actually set up, and which have you been avoiding? Drop it below. ๐