AI skills, tools and MCP: how the parts work together

Start here

Meet skills and MCP

A repeated task can involve several jobs: keeping the instructions, finding the information and arranging it into an answer. Skills and connections may help with different parts of that work.

Two ideas. Tap to explore.

Skill↓ MCP↓

Skill: instructions you can reuse

Imagine you want your workshop summary arranged the same way each week: decisions, tasks and unanswered questions.

A skill can save those instructions so an assistant that supports skills can use the method again.

You still supply the notes and check the result.

MCP: a common way to connect

Imagine asking an AI assistant, “Can you find my workshop notes?” It would need a way to ask a file-search tool for those notes.

MCP stands for Model Context Protocol. It gives the AI app and the program offering that tool a shared way to communicate. Both need to support MCP, rather like devices needing a matching connector.

Illustrations only. No AI or connections run here.

Worked example

A prompt tells the assistant what you want this time

Imagine you help organise a small workshop. You want a short summary of the plan, who is doing what and what is still missing. Here is an example of workshop notes saved in a text file.

Workshop: Saturday, 10am
Mira: bring cards
Room: not confirmed

A prompt is the request you give an assistant. You could paste those notes into a chatbot with this request:

Turn the supplied workshop notes into a brief.
Use three headings: Decisions, Actions, Unknowns.
Keep names and times as supplied.
Put missing details under Unknowns rather than guessing.
Return the draft here for me to review.

The notes do not say which Saturday, which room or when Mira needs to bring the cards. Those questions remain unanswered in the summary, ready for you to follow up.

Skills and tools

A skill keeps instructions; a tool does a job

Suppose you prepare this summary every week. A skill can keep the instructions about headings and missing details together, so an assistant that supports skills can use the method again. You still need to give it the notes, or a way to get them.

A tool lets the assistant do a particular job through its app. For example, a file-reading tool could bring in the notes file instead of you pasting the text. Reading the file and deciding how to arrange the summary are two different jobs.

For occasional work, keeping a copy of your prompt may be enough. A skill may help when you reuse a method often; a tool may help when getting the information is the repetitive part. You can use either on its own. Some tools are built into an app, so using a tool does not always involve MCP.

The connection

How the app asks another program for the notes

Suppose the file-reading tool is offered by a separate program. The AI app needs a way to ask that program for the notes and receive its reply. MCP, short for Model Context Protocol, provides a shared way to make that exchange.

The AI app and the program both need to support MCP. Once they are connected, the app may be able to request an available tool, such as the file reader. MCP describes how they communicate; the tool does the reading, and the skill supplies the instructions for the summary.

What the assistant can reach depends on the tool and the permissions allowed. A connection may offer only reading, or it may also offer actions such as changing a file.

Skill format

What saved instructions look like

What do saved instructions look like? In the Agent Skills format, they live in a file called SKILL.md inside a folder. That file gives the skill a name, explains when it may be useful and describes the method. Supporting templates, reference files or scripts may sit alongside it.

The workshop skill below contains only instructions for arranging the notes into a summary. It does not connect an account or grant access to a file.

---
name: workshop-brief
description: Prepare Decisions, Actions and Unknowns from supplied workshop notes. Use when the user asks for a workshop brief.
---
Use the source notes supplied for this task.
Organize the draft under Decisions, Actions and Unknowns.
Preserve supplied names and times.
List missing dates, rooms and deadlines as unknown.
Identify the source filename when it is available.
Return the brief for review.

Access and actions

Reading notes is different from changing them

Reading the notes file and changing it are different jobs. So are drafting an invitation and sending it. Preparing the workshop summary only needs the notes and a draft for you to review.

Writing a filename in a skill does not give the assistant access to that file. And asking it to read only one file does not necessarily prevent a connection from reaching others. The app and service settings determine the actual limits.

If you later explore a real connection, consider whether your task needs reading, changing something, or both. The provider’s current documentation can help you check the questions below. Programs running on your computer may need supporting software; remote services may need an account or permission to connect.

Check the result

Check the summary against your notes

A clear-looking summary can still use the wrong notes or misunderstand the right ones. You could compare the sample summary with the three lines of notes: are the name and time unchanged, and are the unanswered questions still visible?

Perhaps change Saturday to Sunday and repeat the request in a chatbot. Does the new summary say Sunday? If the room is still unconfirmed, does it stay unanswered? You can try this with pasted notes; no connection is needed. If you later try a different app, these same notes give you a small example whose answer you can check.

If an app shows that it read a file, that tells you about the activity. You would still need to compare the answer with the file to judge whether it understood the contents.

Takeaway

Choose what helps with your task

For a one-off summary, a clear prompt and pasted notes may be enough. If you repeat the same method, a skill may help. If getting the notes is the repeated work, a tool may help. MCP is one way an AI app and a program offering tools can communicate.

Before adding a skill or connection, you could note which repeated job you want help with. After trying it, consider whether it saved enough work to keep using, including the time spent checking and correcting the result.

Task:
What I can already do with a prompt:
Instructions I keep repeating:
Information or action I need:
Skill, tool or connection I am considering:
App and documentation version checked:
What it can reach or do, if known:
What worked in the workshop example:
What I corrected or still do not know:
Would I keep it, remove it or try again later — and why?

Found this useful? Get occasional emails on AI tools worth a look. No spam, unsubscribe anytime.

AI Guide note

How to use this guide

AI Guides are general editorial guidance, not professional advice or guarantees about accuracy, safety, suitability, performance, or outcomes. Tools, terms, prices, features, and laws can change. Check important details against original sources, product terms, reliable references, and qualified help where needed.

Source trail

Formats and connections behind the example

The Agent Skills documentation describes the skill format. The MCP documentation explains how apps and servers communicate. The workshop notes and sample summary are examples.