Skills
Choose your Claude Code surface by asking 'who starts the action' — and put guardrails in hooks
The clean 2026 separation: slash command when a human should explicitly start a workflow, skill when Claude should load a reusable method/checklist on its own, hook when a lifecycle event must run every time. The hard rule: if something MUST happen, it belongs in a hook — hooks are deterministic and Claude can't reinterpret or skip them, whereas prompt instructions are 'polite suggestions.' Start with three hooks (auto-format, sensitive-file protection, notifications) and grow from there.
Source
↳ Follow the thread