Skills, Repos and MCP
What these three actually are, explained without assuming you write code. The widest open gap in the category.
Three words you keep meeting — skill, repo, MCP — explained without assuming you write code. This is the widest open gap in the category: almost everything written about them assumes you are a developer.
A repo, in one sentence
A folder with a memory. It holds files, and it remembers every change anyone ever made to them — who, when, and what exactly changed.
That memory is the whole point. You can always go back. Most AI tooling is distributed this way, which is why "clone the repo" keeps appearing in instructions.
Fuller explanation: Repos, explained.
A skill, in one sentence
A written set of instructions a model loads when it recognises the task.
Not code — prose. A skill for writing legal contracts is a document explaining how your organisation writes contracts, which the model reads before starting. It turns "do this the way we do it" into something reusable.
Fuller explanation: AI skills, explained.
MCP, in one sentence
A standard way to hand a model a tool.
Instead of only producing text, a model connected to an MCP server can call actions: read this folder, query this database, post to this channel. The model still decides what to do; the server decides what is possible.
The permissions you grant an MCP server matter more than the prompt you write. A carefully worded instruction cannot stop an action the server allows, and a badly worded one cannot cause an action the server forbids.
Which is why guardrails is the page to read next.
How they fit together
- A repo is where things live.
- A skill is how the model knows your way of doing something.
- An MCP server is what lets it act rather than only describe.
You can use any one without the others. Most useful setups combine all three — and the moment they do, the permission question stops being theoretical.
What to install, honestly
Start with nothing. Add one thing, use it for a week, then add the next.
The failure mode with all three is the same: installing a stack you do not understand, then being unable to tell whether it is working. A single tool you can reason about beats six you cannot.