Agent Mentor, taken apart

For readers who want the full picture: what it is made of, what actually happens when your agent writes a lesson set, how quality is enforced, and why the boundaries sit where they do. This page goes deep on producing a lesson set; review, update, and export live in the loop on the home page.

You install a folder of one skill set, all plain text and auditable

Four legs plus one intent router; every part is plain text your agent can read and you can edit. Each path below exists in the repository.

├─ SKILL.md

├─ agent-mentor/skills/generate-course-from-topic

├─ agent-mentor/skills/review-course

├─ agent-mentor/skills/maintain-course

├─ agent-mentor/skills/export-course

├─ scripts

└─ course-reader

The top-level intent router: one sentence goes to produce, review, update, or export.

Produce leg: the full contract for asking, researching, writing lessons, and checking.

Review leg: works out what's due on a Leitner schedule, then quizzes you, answer first, explanation after.

Maintain leg: change difficulty, add lessons, health-check the library; it snapshots before editing.

Export leg: the whole lesson set to HTML/PDF, the glossary to Anki cards.

Deterministic scripts: course-guard checks, review-mark logs reviews, course-doctor audits, course-snapshot backs up.

The local reader: rendering, practice, the review page, and the export core.

One lesson set, dissected

Five stages. Each has rules the skill hard-codes, and each leaves an inspectable artifact.

01
Request
Completeness is judged first: a topic with no background gets questions, not lessons.
The original request is stored in agentmentor.json
Create the English showcase course learn-use-agent-mentor teaching buyers how to operate the Agent Mentor Skill with their own coding agent — requesting courses, practicing in the reader, running review sessions, maintaining courses, and exporting them — aimed at non-developer buyers and intended to lead the course gallery.
02
Learner model
Known concepts, unknown concepts, and the first thing the learner should manage alone.
learnerBrief and learnerModel
A beginner English course for someone who has a coding agent (Claude Code, Codex, or similar) with the Agent Mentor Skill installed but has not used it for structured learning. The course teaches the five operating requests as plain-language sentences to the agent: ask for a course, work through a lesson, run a review session, update a course, export a course. No terminal, IDE, or programming knowledge is assumed; the learner never types a command.
03
Research
The host agent's search finds authoritative sources; marketing copy and pure AI output fail the authority check.
sources.md
04
Writing
Lessons follow the contract: bounded new concepts per lesson, every interaction serves one understanding point.
Lesson Markdown and glossary.json
05
Checks and rendering
course-guard runs; only a full pass starts the local site.
Guard status and the local site

Quality system: what fails does not ship

course-guard is a set of deterministic checks — twenty-plus of them, in four groups, no model calls.

Sources and tracing

Lesson citations must be complete; every glossary term needs an origin; a source list that is all marketing or AI-generated fails outright.

Structure and format

The README schedule must cover every lesson; prev/next links, template sections, and frontmatter are all mandatory.

Practice and interaction

Every block type is validated: feedback must explain the mechanism, not just say right or wrong; mermaid diagrams need surrounding prose.

Difficulty and learner fit

Non-code line counts are capped to prevent concept overload; the learner model, request completeness, and acceptance scores are checked together.

Sample: course-guard passed

Eight practice blocks, one understanding each

Interaction is not decoration. Each block checks a single understanding point, and feedback explains the mechanism.

Choose

Tell close options apart at a concept boundary.

Order

Arrange a flow or dependency into the right sequence.

Predict output

Guess what the code prints, then verify.

Trace variables

Simulate execution by hand, one step at a time.

Fill the slots

Complete only the key decisions — no rote rewriting.

Fix the bug

Find and change the one broken line in working code.

Read the diff

Judge whether a change really alters behavior.

Hotspot

Point at where the problem happens on a diagram.

Try one live

Fill the two blanks so the function returns the largest value in the list.

A fill-in exercise leaves only the key decision points, checking whether you know what to reach for instead of making you write it all from scratch.

def largest(nums):
    best = nums[0]
    for n in nums:
        if :
            best = 
    return best

Get access

Paste once. Updates follow.

After payment you automatically receive an install prompt. Paste it to your agent and the installation is done. Automatic delivery with no waiting on a human, and no cloud account. Reference price $29.90. The platform page has the final price.

Get it on Paddle

Already purchased? See the install guide →

I bought Agent Mentor. Please install and start it: download its zip from https://●●●●●●/download?k=●●●●●●, unzip it into your skill directory, save the download URL and key to ~/.agent-mentor/update.conf, then start the reader and tell me how to make my first lesson set.

A sample of the install prompt. The full version arrives automatically after purchase; paste it to your agent.