Quick Reference: Skills, Plugins & Connectors

Key distinction: Skills cannot make network calls. Connectors cannot apply procedural logic. You need both for most real-world workflows.

• Authentication: One-time browser-based OAuth flow — Claude never sees your password
• Permissions: Granular and revocable — read-only vs read-write, specific folders vs entire account
• Token storage: Tied to your authenticated Cowork session, not shareable across devices
• Enterprise provisioning: Admins can pre-install and configure connectors for all users centrally

Common traps:

• Installing a connector does not complete authentication — the browser OAuth step is separate
• A connector does not see all company data — only what was explicitly granted during OAuth
• MCP Apps work across multiple AI clients (VS Code, etc.), not just Claude

Built-in Agent Skills (no installation needed):

• Excel/CSV reading, writing, formula creation
• Word document generation with formatting
• PowerPoint presentation creation
• PDF reading and analysis

Custom Skills (SKILL.md files):

• Written in plain-English Markdown, not code
• Stored in ~/.claude/skills/ (the only discoverable path)
• Must include YAML frontmatter (name, version, triggers) — without it, the skill is invisible
• Created via /skill-creator wizard or manually
• Cannot make network calls — need a Connector for external data

When to build custom: When your organisation has proprietary procedures, templates, or scoring criteria that no marketplace plugin covers.

---
name: "Meeting Summariser"
version: "1.0"
description: "Extracts action items from meeting notes"
triggers:
  - "when the user uploads meeting notes"
  - "when the user says /summarise-meeting"
---

## Instructions
1. Read the uploaded meeting notes
2. Extract all action items with owners and deadlines
3. Format as a structured summary using the template below

## Output Template
**Meeting:** [Title]
**Date:** [Date]
**Action Items:**
- [ ] [Action] — Owner: [Name] — Due: [Date]

Critical rules:

• YAML frontmatter between --- markers is mandatory
• Store in ~/.claude/skills/ or Claude cannot find it
• Trigger conditions must be specific (not "when the user asks about documents")

The standard pattern: Connector (fetch data) → Claude reasoning (process) → Skill (format output)

Tool priority hierarchy:

1. Connectors — structured API data, fastest and most reliable
2. Browser — web navigation, slower but works without connectors
3. Computer Use — screen interaction, slowest but handles legacy apps

Design rules:

• Each component should have a single responsibility
• Chain smaller, specialised skills rather than one monolithic skill
• Minimise active connectors — unused tool definitions still consume context tokens
• Use MCP tool-call batching for simultaneous data requests

Common trap: Over-engineering with 10 connectors when 2-3 would suffice. Each adds latency, token overhead, and failure risk.