$ learn
What is the difference between Skills, MCP, Connectors and Plugins?
Understand how skills, connectors, MCP and plugins work together through an email follow-up experiment.
The short version
When you ask AI to follow messages from an organization, it needs a way to reach your email and instructions explaining how you want the results organized. Those are different things.
The episode uses a counter clerk to explain a skill, a connector, MCP and a plugin. We then try the same request twice with the connector enabled both times: once without the skill and once with it.
Key terms
- Skill: reusable instructions describing how to perform a task.
- Connector: a way to link an AI application to a specific service, such as email or files.
- MCP: an open standard defining how AI applications communicate with tools and services that support it.
- Plugin: a package for a particular job that may combine skills, connectors and other components, depending on the application.
The explanation
The clerk and the procedure binder
Imagine the chat as a counter and the model as the clerk behind it. Before you connect email, the clerk cannot see your messages just because you ask. Even after connecting, it does not automatically know how you prefer them organized.
A skill is like a procedure binder on a shelf: you write down the method and boundaries, and the application can use those instructions when they fit the task. Do not assume it will always select or follow them; check what the application used and inspect the result. A skill can help when you find yourself repeating the same instructions.
In the episode, the sender-tracking skill calls for sorting messages into four groups, showing dates and costs, ordering them by the nearest date and leaving space for the account owner's decision. It says to read only, without replying or deleting.
The connector and MCP
A connector is like a line from the clerk to the email office. MCP is closer to a shared communication specification; it is not the email account itself. Different services and applications can support the standard, but that does not mean every connector uses MCP or that every application supports every service in the same way.
In the recording, the Gmail consent screen lists permissions including reading messages, writing drafts and sending email. Permissions and interfaces vary by service and account; read the actual screen before agreeing.
A plugin is a kit
A plugin is like a ready-made kit for a particular job: it may contain instruction binders and connectors together. In the episode's example, the plugin page shows one skill and two connectors. Contents and installation differ across applications and plugins.
What changed in the experiment?
The connector was enabled both times. Without the skill, Claude found and sorted the emails in its own way, without dates beside the items. With the skill, it opened the instructions and returned a summary identifying one message needing a decision, with webinar dates, a marker for a passed date and space for the decision. This is what happened in the recording, not a guarantee of your result.
Try it yourself
Choose an organization that emails you often. Review the email connection's permissions before agreeing, and connect a work or university account only if you are authorized. Connection steps and availability may differ from the episode.
In a new conversation, ask to follow the organization's messages, replacing its name in the prompt. Check the original emails against the summary.
Follow the emails from [organization name], summarize the important items and tell me whether anything needs my decision. Read only; do not reply to or delete any message.For the second attempt, add the following instructions to your request or use them as a starting point for a skill in an application that supports one. This starter is based on the instructions shown in the episode; it is not a copy of the original file. Define the four groups to suit your needs first.
Follow messages from [organization name] and sort them into four groups: [list the groups]. Include dates and costs when present, order items by nearest date, mark dates that have passed and leave a space for my decision. Read only; do not reply, send or delete any message. If information is missing, say it is not stated.Compare the two results and check dates, amounts and original messages. Notice whether the instructions were actually used, and revise them to suit your work.
Why this matters for building software
When building an AI-assisted workflow, separating instructions from service connections helps you identify what needs changing: the way a task is performed or where its information comes from. You remain responsible for reviewing access and results.
Related lessons
- What is an AI agent?: how a model uses tools to perform a task.
- How do you ask for exactly what you want?: describing the task and desired result.
- Why does AI lie so confidently?: why results need checking.