Understanding the Ecosystem

What You Need to Know

Cowork ships with a three-part ecosystem that determines how Claude interacts with external tools, follows your procedures, and chains capabilities together. Get the mental model right and everything else in Domain 3 falls into place. Get it wrong and you'll misdiagnose every integration problem, install the wrong components, and waste hours troubleshooting workflows that were never going to work.

The three components are Skills, Connectors, and Plugins. Each has a specific job. None of them can do the other's job. The exam will test whether you can tell them apart under pressure — and most candidates get it wrong at least once.

Skills: Claude's internal recipe book

A Skill is a set of instructions that tells Claude how to do a specific job. Think of it as a page in an employee handbook — it describes the steps, the expected output format, edge cases to watch for, and examples of what good results look like. Skills are written in plain Markdown, stored locally on your machine, and loaded into Claude's context only when needed.

The critical point: Skills cannot reach outside Claude's environment. A Skill can't make a network call, query a database, or pull data from Salesforce. It handles the "how" — how to format a report, how to apply your brand voice, how to score a legal contract against your 15-point checklist. If your workflow needs data from an external service, the Skill isn't the component that fetches it.

Connectors: the bridge to external services

A Connector does exactly one thing: it gives Claude access to an external service. Google Drive, Slack, Salesforce, Jira, a SQL database — each needs its own Connector. The Connector handles authentication (via OAuth, so Claude never sees your password), manages permissions (read-only vs. read-write), and provides a structured interface for fetching or pushing data.

Connectors handle the "where from" and "where to." They bring data into Claude's environment so a Skill can process it, or they push results out to an external destination after processing is complete.

Plugins: the complete package

A Plugin is a distribution bundle — the installable package you find in the Browse Plugins catalogue. A single Plugin can contain one or more Skills, one or more Connectors, and slash commands for quick access. When you install a "Salesforce Plugin," you get the Connector for API access, the Skills for data analysis and formatting, and the commands for common actions — all in one go.

Think of Plugins as meal kits. The Connector is the delivery of fresh ingredients, the Skill is the recipe that tells you what to do with them, and the slash commands are the pre-measured spice packets for common additions. You install the kit; the components work together.

MCP: the universal connector standard

The Model Context Protocol (MCP) is the open standard that makes Connectors work. Before MCP, every tool integration needed custom code — a bespoke adapter for Slack, another for Jira, another for Google Drive. MCP provides a universal interface so any tool can plug into Claude (or any other AI client) without spaghetti integrations.

MCP is governed by the Linux Foundation's Agentic AI Foundation, not proprietary to Anthropic. This matters for the exam: any question claiming MCP is Anthropic-exclusive technology is wrong. The protocol is open, cross-platform, and used by multiple AI clients beyond Claude.

Progressive disclosure: how Claude manages the token budget

Claude doesn't read your entire skill library at the start of every session. That would waste thousands of tokens on capabilities you may never use. Instead, it employs progressive disclosure: it scans metadata labels first — the skill's name, description, and trigger conditions in the YAML frontmatter — and only loads the full instructions when it is confident a specific skill is needed for the current task.

Think of it as scanning book spines on a shelf rather than reading every book cover to cover. This isn't just an efficiency feature. It's the reason you can have dozens of skills installed without degrading Claude's performance. The exam tests this: if a question asks why having many skills doesn't slow Claude down, the answer is progressive disclosure.

The data flow architecture

When Skills and Connectors work together, data flows through a clear pipeline:

1. Connector fetches — raw data arrives from an external service (Salesforce figures, Slack messages, Google Drive documents)
2. Claude reasons — the core model analyses, synthesises, and makes decisions about the data
3. Skill applies — specialised procedural logic formats, scores, or transforms the output according to your rules
4. Connector pushes (optional) — the finished result goes to an external destination (email, Slack channel, shared drive)

Each component has a single responsibility. The Connector doesn't analyse. The Skill doesn't fetch. Claude's reasoning engine sits in the middle, orchestrating the flow. When you design workflows (covered in depth in Task 3.5), keeping this separation clean is what makes them reliable and maintainable.

---

Common Mistakes

---

Hands-On Activity

---

Practice Question

---

Sources

• Use plugins in Cowork — Anthropic
• Cowork: Claude Code power for knowledge work — Anthropic
• Claude Cowork Guide 2026: Skills, Plugins, Connectors — FindSkill.ai