# Day 4 — Claude Skills, Subagents & Automation

- A Skill is one reusable capability stored in one `SKILL.md`: name and trigger description, allowed tools, ordered instructions, and a quality standard.
- Narrow skills fire reliably. A skill that tries to cover several unrelated jobs, assumes hidden context, or has no quality bar will be inconsistent.
- Slash commands call a skill and produce the same output shape each time.
- Automation has four building blocks: trigger, condition, action, and optional delay. A webhook is an instant event notification; an API is the communication channel.
- MCP (Model Context Protocol) gives Claude scoped access to Drive, Gmail, Slack, Notion, Calendar, and CRMs.
- Treat connections like permissions. Connect only what the task needs, never expose raw credentials, and ask before adding a client connection.
- A subagent can run automatically or on request. An Agent Team is a manually invoked group of agents and is token-heavy. A Workflow is a manually invoked, token-heavy path requiring Max.
- A Routine adds a schedule so a prompt or skill can run hourly, daily, or weekly without a person pressing go.
- Good first builds include lead dossiers, department status summaries, meeting-to-recap drafts, and Morning Briefs.
- Zapier and n8n remain useful for always-on, deterministic plumbing. Claude should handle judgment; automation tools should move data on reliable rails.
- Automation must include error handling, idempotency, human checkpoints for client or financial actions, narrow permissions, and a test run before real data.
- Shortwave demonstrates the same closed-loop pattern for inbox work: draft, attach context, leave a note, and require a human send checkpoint.
- Homework: build and test a narrowly scoped skill, connect one real tool, and explain the result in a short Loom.

## Relevance to the final project

- The final funnel needs reusable SOPs and workflow logic, with human review before any client-facing or irreversible action.