Lesson 01: Ask for a course

Lesson objectives:

• Describe what happens between your request and the course link your agent hands back.
• Judge whether a course request will trigger clarifying questions or go straight to generation.
• Write a course request that names the topic, your starting point, your gaps, and your goal.

Prerequisites: None | Previous << README | Next 02 Work through a lesson >>

The skill is installed — so where is the on switch?

There is no app icon, no dashboard, no login. That absence is the design, and it is the first thing to understand. A skill is a folder of instruction files that your coding agent reads when a request matches — the everyday handle is "a playbook your agent follows." Agent Mentor's playbook activates on any learning request: "I want to learn X," "teach me X," "make me a course on X" all route into it, whether or not you mention the product by name¹. The name Agent Mentor Skill exists so you can refer to it; it is not a trigger phrase you have to remember².

So the on switch is a sentence. What that sentence should contain is this lesson's subject, because the sentence decides whether you get a course built for you or a course built for a stranger.

Explanation

Start with what your agent does after you send a learning request. It reads the skill's instructions, checks whether your request is clear enough, gathers sources it can cite, writes the course as files on your machine, runs a mechanical quality check on those files, starts a local course site, and hands you a link ending in /courses². The whole chain looks like this:

What this diagram wants you to see is where the course lives: not on a website somewhere, but as files next to your agent, rendered by a local reader, with a path back to the agent when you get stuck. Directing an agent through structured work is itself a taught subject now — the model makers run beginner courses on it³ — and this skill packages one loop of it: the agent does the building, and you supply the two things it cannot guess⁴.

Those two things are your starting point and your goal. The skill sorts requests into rough tiers². A topic-only request — "I want to learn Next.js" — forces the agent to stop and ask one to three short questions, because it cannot set the course's difficulty without knowing what you already know. A request that names the topic, your starting point, and your gaps skips the questions and goes straight to generation. You do not need to write a specification; four short parts are enough:

1. Topic — what you want to learn.
2. Starting point — what you already know or use.
3. Gaps — what you do not know, named plainly.
4. Goal — one thing you want to be able to do at the end.

The gaps part carries a guarantee: anything you flag as unknown may not be treated as a prerequisite. The skill records it in the course's learner profile, and the course must bridge it instead of assuming it⁵. Saying "I don't know X" does not make the course worse; it makes the first lesson land where you are².

One more thing happens on your very first request: setup. The skill installs its own copy of the tools it needs into a folder in your home directory, installs the reader's dependencies, and starts the reader — reporting progress as it goes². The security boundary is worth stating outright: everything runs in your user space with no administrator password, the reader listens only on your own machine, and the one network call on later runs is a version check that carries no user data². Deleting that one folder undoes the install. When the agent hands you the link, use the full /courses address it gives you — that page is your course list; the bare domain is a product page².

Worked example (follow along)

Rosa does marketing reports and wants to learn pivot tables. Watch the same wish written twice.

First draft: "I want to learn pivot tables."

Topic only. The agent must reply with questions — roughly: what do you use spreadsheets for today, have you used any pivot table before, what do you want to produce? Three messages later, the course starts. Nothing went wrong, but the questions were predictable, and Rosa could have answered them up front.

Second draft: "I want to learn pivot tables in Google Sheets. I build a monthly report by copying totals around by hand, and I've never inserted a pivot table. By the end I want to turn our sales sheet into a by-region summary on my own."

Check it against the four parts: topic (pivot tables in Google Sheets), starting point (builds reports by hand), gap (never inserted one), goal (a by-region summary, unassisted). The agent asks nothing, builds the course, runs its quality check, and replies with a /courses link. The goal sentence quietly does a second job: the course's final lesson now has a target — Rosa's own sales sheet — instead of a generic dataset.

Your turn (faded example)

Complete this request for a topic of your own. The structure is done; the judgment calls are yours.

"I want to learn ____. I already ____ (what you use or do today), but I've never ____ (the gap, stated plainly). By the end I want to ____ (one thing you'll do without help)."

Answer: Any filling works if it survives two tests: a stranger reading only your sentence could guess what lesson 1 should assume, and the goal names an action — something you will do — not a feeling of understanding. "By the end I want to feel confident" fails the second test; "by the end I want to write a formula that flags overdue rows" passes.

Summary + what's next

You can now describe the chain from request to local link, and you can write the four-part request that skips the question loop: topic, starting point, gaps, goal. The gap you name becomes a promise the course has to keep. Next lesson: the link is open, the page looks like an article — and reading it like an article wastes most of what is on it.