Separate skills from prompts, rules, tools, and MCP

A skill is a folder with instructions an agent loads when a task matches. It teaches the agent how to do a class of work: the steps, the checks, the things that go wrong.

That sounds close to a prompt or a rules file. It isn’t.

Five things that look alike

A prompt asks for one result. “Review this repo before I release it.” Once answered, it’s gone.

A rule applies every time the agent works in a project. It lives in a file like AGENTS.md and says things like “run npm test before you call anything done”. I wrote a full guide on what goes in an AGENTS.md file.

A tool gives the agent a capability it doesn’t have on its own: read a file, run a command, call an API.

MCP (Model Context Protocol) is one standard way to connect tools and data sources to an agent. It’s plumbing. It doesn’t say how to do the job.

A skill is the how. It’s a procedure the agent discovers when the task fits, loads, and follows.

The agent is the loop that ties it all together: pick an action, look at the result, pick the next one.

How they work together

Take the project we build in this course, a release-readiness skill:

• your prompt names the release to review
• the project rules forbid force-pushing and name the test command
• the shell and Git tools let the agent inspect the repo
• MCP might connect it to your issue tracker
• the skill says which checks to run, in which order, and how to write the report

Take one piece out and something breaks. Without the skill, the agent improvises a different checklist every time.

When not to write a skill

Don’t turn every useful paragraph into a skill. My rule of thumb:

• a fact about the project goes in the docs or in AGENTS.md
• a permanent safety boundary goes in rules or permissions, because an instruction can be ignored and a permission can’t
• a remote operation needs a tool
• a repeatable procedure the agent should find on its own, that’s a skill

Quick test: if you’d repeat the same five-step explanation to a new coworker every time, that’s a skill. If you’d say it once on day one, that’s a rule.

Try this on the release project: write five cards named Prompt, Rule, Skill, Tool, and MCP. Put one release-readiness responsibility on each. If two cards claim the same job, decide which layer owns it before you go on.

New to context windows and tokens? Start with the AI fundamentals course. Skills make more sense after it.