Sources

S1 — Agent Mentor Skill package entrypoint (intent router)

URL: SKILL.md

  • authority: official-docs
  • supports: The skill activates on any learning request without a trigger phrase, and routes review ("quiz me"), maintenance ("update my course"), and export ("export this course") requests to their own workflows.
  • key-fact: "Use for EVERY learning request, even when the skill is not named — 'I want to learn X', 'teach me X'... Also handles review (quiz me), maintenance (update the course), and export."

S2 — Course-generation workflow contract

URL: agent-mentor/skills/generate-course-from-topic/SKILL.md

  • authority: official-docs
  • supports: The user-facing product name; the request tiers (topic-only requests trigger 1-3 clarifying questions, requests with starting point and gaps go straight to generation); unknown concepts must not be smuggled in as prerequisites; the first-run environment setup is user-space with no sudo and collects no user data; the reader listens on localhost only and the URL handed to the user must end in /courses; the update check sends only a version query.
  • key-fact: "Too vague: 'I want to learn Next.js' gives only a topic name. Stop and ask 1-3 short questions first... When the user says 'I don't know X,' X must go into unknownConcepts and must never be smuggled in as a prerequisite in lesson 1... Stage 0 runs entirely in user space, no sudo, and collects or exfiltrates no user data... it must include the /courses path."

S3 — Course authoring guide (course shape and quality gate)

URL: agent-mentor/skills/generate-course-from-topic/course-authoring-guide.md

  • authority: official-docs
  • supports: Every lesson follows a fixed skeleton (objectives, explanation, worked example, faded example, exercises, summary); key facts must carry citations into sources.md and cannot be fabricated; every course ships a glossary with sourced definitions; exercises carry completion criteria for self-judgment; interaction feedback must explain the mechanism, never just say wrong; the mechanical quality check (course-guard) gates delivery and maintenance.
  • key-fact: "Every key fact/API/version/definition that makes it into the course must be corroborated... every question has the four-part structure: approach guidance, quantifiable completion criteria, the answer or key points, progressive hints... every error-feedback message must explain the mechanism, never just say 'wrong.'"

S4 — Review-course workflow contract

URL: agent-mentor/skills/review-course/SKILL.md

  • authority: official-docs
  • supports: The review triggers; the pool contents (glossary terms once a lesson of the course is complete, interaction blocks only after being answered); Leitner boxes 1-5 with intervals 1/3/7/14/30 days and due computation; at most 10 items per session with previously-wrong items first; quiz-first-explain-after for terms, rephrase-plus-variant for blocks; "I don't know / I forgot" recorded as wrong with a non-judgmental tone; immediate per-item recording; the closing report with tomorrow's due count; the zero-due and no-progress branches.
  • key-fact: "Boxes 1-5 map to intervals [1, 3, 7, 14, 30] days... Take at most 10 items per review session... quiz first, explain after... Treat 'I don't know / I forgot' as a wrong answer; keep the tone non-judgmental... report: 'Reviewed X items today, Y correct, Z wrong, N items due tomorrow'... If there are 0 due items... do not manufacture questions."

S5 — Maintain-course workflow contract

URL: agent-mentor/skills/maintain-course/SKILL.md

  • authority: official-docs
  • supports: Maintenance triggers; the announce → snapshot → edit → verify → report sequence with quality-never-regresses and rollback; "undo that" restores the latest snapshot and re-runs the check; review records are keyed by lesson filename, block id, and term text, so renames orphan them and require an explicit warning; the library health check is read-only; extension lessons are capped at depth 1 and three per parent lesson.
  • key-fact: "The edit MUST pass course-guard.mjs after completion (quality NEVER regresses), every edit is reversible (automatic hidden snapshot), and edits NEVER silently break the review loop (orphaned review records are surfaced explicitly)... explicitly tell the user: 'This will orphan roughly N review records'... Depth capped at 1... once a parent already has 3 extensions, refuse to generate a 4th."

S6 — Export-course workflow contract

URL: agent-mentor/skills/export-course/SKILL.md

  • authority: official-docs
  • supports: Export triggers and format mapping (both formats by default, HTML for print/PDF requests, CSV for flashcard requests); output lands in the repo's exports folder per course and the path is reported; PDF comes from the browser's print function; Anki import via File → Import with front/back field mapping; export is fully local, only writes to exports, and is a snapshot requiring re-export after changes.
  • key-fact: "If both are mentioned, or they just say 'export this course' without specifying → export both... for a PDF, use the browser's 'Print → Save as PDF'... open Anki → File → Import → select the -cards.csv file, map the front/back fields... The export is a snapshot of the course's current content; re-export after any course change."

S7 — Anthropic: Building Effective AI Agents

URL: https://www.anthropic.com/engineering/building-effective-agents

  • authority: official-docs
  • supports: The framing that an agent directs its own process and tool use to do work — the reason the learner's job in this course is supplying goals and checking results rather than performing the steps.
  • key-fact: "Agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."

S8 — Wikipedia: Leitner system

URL: https://en.wikipedia.org/wiki/Leitner_system

  • authority: encyclopedia
  • supports: The general mechanism behind the review schedule: correctly answered items advance to boxes reviewed at longer intervals, incorrectly answered items return to the frequent-review end.
  • key-fact: "Correctly answered cards are advanced to the next, less frequent box, while incorrectly answered cards return to the first box for more aggressive review and repetition."

S9 — Learner model contract

URL: docs/learner-model-contract.md

  • authority: official-docs
  • supports: The guarantee attached to naming your gaps: a concept the user says they don't know goes into the course's learner profile as an unknown concept and must not be used as a first-step prerequisite; the course must bridge it.
  • key-fact: "If the user says they do not know X, put X in unknownConcepts; do not use X as a first-step prerequisite... The first independent action must be executable using only knownConcepts plus information introduced in the same lesson."

S10 — Reader drill-down design: three copy-to-agent entry points

URL: docs/superpowers/specs/2026-07-10-course-drilldown-design.md

  • authority: official-docs
  • supports: The selection chip ("ask agent about this passage") copies the lesson path plus the selected text verbatim; the glossary popover carries a "have the agent teach this" action with the term, definition, and lesson path; the lesson-end "go deeper" entry copies the lesson path with a deepening request; the shared triage rule answers small questions in chat and grows at most one extension lesson for real gaps.
  • key-fact: "Copied content = lesson path + the selected text quoted verbatim + triage instructions... rules for the agent: read the lesson file and glossary first; if one or two conversational turns can explain it, explain directly and generate no files; only a real foundational gap or systematic deepening need generates one extension lesson."

S11 — When-stuck help card design

URL: docs/superpowers/specs/2026-07-10-stuck-help-card-design.md

  • authority: official-docs
  • supports: The course page carries a when-stuck card with situation-specific templates (concept unclear, code error, exercise without a plan) plus a lesson-end reminder; templates auto-fill the course location, force the learner to fill in where they are stuck and what they tried, and instruct the agent to lead with hints and questions rather than the full answer.
  • key-fact: "The unified skeleton: 'I am learning __ (auto-filled) + situation-specific blanks + behavioral constraints for the agent'... do not give the complete explanation directly: first explain with an analogy, then ask a question to confirm my understanding."

S12 — Anthropic Academy

URL: https://www.anthropic.com/learn

  • authority: official-docs
  • supports: Directing an agent through structured work is now taught as its own subject by the model makers, which grounds lesson 1's framing that operating an agent is a learnable, established practice rather than a niche trick.
  • key-fact: "Featured courses include AI Fluency tracks for non-developers alongside Claude Code and API tracks."