Agent Mentor: from first request to export · Lesson 4 of 5

Lesson 04: Update a course

Lesson objectives:

  • Phrase the four kinds of maintenance request: fix content, adjust difficulty, add material, health-check the library.
  • Name the guardrails the agent must follow around any edit: announce, snapshot, verify, and undo.
  • Judge which changes put review records at risk and what warning to expect first.

Prerequisites: A generated course of your own | Previous << 03 Run a review session | Next 05 Export a course >>

Your course is a set of files, and files drift

A fact in lesson 3 goes stale. Lesson 2 turned out harder than your starting point. You wish there were one more lesson on the part you now care about. With a printed book you'd live with it; your course is files on your machine, and the same agent that wrote them can rewrite them. What keeps this from being scary — an AI editing your study materials in place — is that the skill binds every edit to guardrails you can watch for and a one-sentence undo.

Explanation

Maintenance has its own trigger phrases, all plain: "update my course," "fix the outdated parts," "this lesson is too hard, simplify it," "add a lesson on X," "health-check my course library"1. Asking to learn a new topic is not maintenance — that routes back to course generation2.

Whatever you ask for, the skill holds the agent to one editing sequence1:

What this diagram wants you to see is that the two boxes around the edit are not optional courtesies. The snapshot comes before any file is touched, which is what makes "undo that" a real sentence later — it restores the most recent snapshot, and the quality check runs again after the restore1. The check comes after, and it is the same mechanical gate a new course must pass: structure, citations, sources, glossary3. An edit that fails it gets fixed or rolled back; the skill's rule is that quality never regresses1. Nothing in this flow deletes anything permanently, and you can watch for the announce and report steps in the agent's replies.

One risk needs your judgment, because the agent can only warn about it, not remove it. Lesson 03's review records are keyed to names inside the course: a lesson's filename, an interaction block's id, a glossary term's exact text1. Rewrite a paragraph and the records don't care. Rename or delete one of those keys, and every record pointing at it is orphaned — it points at nothing, and that slice of your review schedule dies with it. The skill requires the agent to warn you first — "this will orphan roughly N review records" — and to prefer edits that keep the keys stable1.

The last request type is the widest: "health-check my course library." The agent runs a read-only scan and reports which courses pass the quality check, which have orphaned records, and which are clean — it changes nothing on its own; fixing anything is a follow-up request you choose1.

And one entry point from lesson 02 quietly belongs to maintenance: the "go deeper on this lesson" link. When a deep-dive is worth more than a chat answer, the agent grows an extension lesson attached below the parent — capped at one level deep and three extensions per lesson, because a lesson needing a fourth patch is a lesson that should be rewritten1.

Worked example (follow along)

A full exchange, guardrails annotated.

You: "Lesson 2 of my pivot-tables course is too hard. I got lost where it assumes I can write formulas."

Agent: "I'll rework lesson 2 to introduce the summary step without formulas, keeping the lesson file, block ids, and terms unchanged." — the announcement, one sentence, and note the promise about keys: no rename, so no records at risk.

Agent: "Snapshot saved. Editing… done. Quality check passes. Lesson 2 now builds the by-region summary through menu steps first; formulas moved to an optional aside. Say 'undo that' if you want it back." — snapshot before, check after, undo offered.

You: "Undo that." (Suppose you preferred the original.)

Agent: "Restored the snapshot from before the edit; quality check passes again." — the backup was real, and the gate ran after the restore too.

The whole transaction is reversible, verified on both ends, and you never opened a file.

Your turn (faded example)

Complete this maintenance request so the agent can act without a follow-up question — and so you've already thought about the keys:

"In my ____ course, lesson ____ is ____ (too hard / outdated / missing something). Change ____, and keep ____ unchanged."

Answer: A filled example — "In my pivot-tables course, lesson 4 is outdated: it shows the old menu layout. Update the menu steps and the screenshots description, and keep the lesson filenames, block ids, and glossary terms unchanged." The last blank is the one that shows lesson-04 thinking: you are telling the agent which keys your review schedule depends on.

Summary + what's next

You can now request all four maintenance moves, and you know the sequence that must wrap any edit: announce, snapshot, edit, verify, report — with "undo that" held in reserve. You track record safety by keys, not edit size, and you expect a warning before any key changes. One operation remains: the course lives next to your agent, but sometimes it has to travel — to paper, to a friend, to a flashcard app. That is the last lesson.

Footnotes

  1. Maintain-course workflow contract ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  2. Agent Mentor Skill package entrypoint (intent router) ↩

  3. Course authoring guide (course shape and quality gate) ↩

Exercises

01

Say "health-check my course library" to your agent. Restate the report in your own words: which courses pass the quality check, whether any orphaned records exist, and which single fix you would do first.

Level 1 (warm-up)
Done criteria · checked locally
02

Request one small content fix on a course you own — a stale fact, a clunky example. Watch for the announce step and the verify step in the agent's replies. Then say "undo that" and confirm the course reads as before.

Level 2 (advanced)
Done criteria · checked locally