Written in full, but still a draft. The depth lives in the v1 pages linked below; this page is the Module 1 version.
Fundamentals: context, MCPs and skills
Three ideas change how well an AI agent works for you. You will use all three on the Hotels repo this week.
Context management
The idea. An agent can only use what is in front of it right now: your messages, the files it has read, and any instruction files it loads. That space is limited. Too little and it guesses. Too much, or the wrong thing, and it gets confused or follows bad advice. That last case is called context poisoning.
What the repo does about it. AGENTS.md is the one instruction file, kept short on purpose. The agent reads it first, so the rules that matter are always in front of it.
Try it (15 minutes).
- Start a fresh agent session in the repo.
- Ask: "How do I run the tests, and what must pass before I commit?"
- Check its answer against
AGENTS.md. - Ask something
AGENTS.mddoes not answer, such as "Who owns the hotel list page?" Notice what it does when it does not know.
You should see: a correct answer to the first question that matches AGENTS.md. For the second, a guess or a request for more information.
Think about it. If AGENTS.md grew to 500 lines, what would go wrong? What would you cut first?
Go deeper: Context management (v1).
MCPs
The idea. An MCP server gives your agent extra abilities, such as reading a ticket system or searching documents. Each one also adds descriptions of its tools to the agent's context, whether or not you use them. More is not better.
Try it (15 minutes).
- List the MCP servers your agent has enabled.
- For each, decide whether you need it to work on the Hotels repo.
- Turn off the ones you do not need for this session.
- Ask the agent what tools it has available, before and after.
You should see: a shorter tool list after you turn servers off, and the same repo tasks still working.
Think about it. What is the cost of leaving a tool on "just in case"? When is it worth paying?
Go deeper: Set up MCP and skills (v1).
Skills
The idea. A skill is a reusable set of instructions you write once and the agent follows whenever it applies. It is how you stop repeating yourself, and how you write a good habit down so it happens every time.
Try it (30 minutes). Write a skill called run-the-gates:
- In plain words, tell the agent to run
./mvnw verifyand thenpre-commit run --all-files. - Tell it to report the result in at most five lines: what passed, what failed, and the one next step.
- Tell it never to change code to make a gate pass unless you ask.
- Save it the way your agent expects, then use it on your Exercise 5 change.
You should see: the agent runs both checks and gives you a short summary instead of pasting a wall of output.
Think about it. The five-line limit is on purpose. Anything you have to approve should be short enough to review. What else would you cap?
Go deeper: Extending agents with skills (v1) and Build your first skill (v1).
TODO
- TODO: confirm the skill format and MCP steps for the agent the cohort uses; this page is deliberately tool-neutral
- TODO: decide how much of the v1 context, MCP and skills material stays as pre-reading