Kit de Leçon IA pour Enseignants

What's in this pack

This is the stack for a working teacher — not the 30-tab "AI for Education" Pinterest board. Every pick here earns a slot in a real week: Monday you plan, Tuesday-Thursday you teach, Friday you grade, and on the side you keep a tutor available so the kid who missed Wednesday isn't stuck.

The pack covers five jobs in order: plan the lesson, generate the materials, write the formative quiz, host it in an LMS your school already approves, and stand up a personalized tutor with memory so every student gets follow-up that knows what they already learned. Everything is open-source or has a clear free tier. Nothing here ships student data to an opaque vendor without you opting in.

Install in this order (plan → material → quiz → grade → tutor)

1. Fabric — start here. It is a library of 100+ AI prompt patterns (summarize, extract_wisdom, create_quiz, write_essay). One CLI, one consistent shape. Use extract_wisdom on a textbook chapter to get a clean outline before you plan a lesson around it.
2. Mr. Ranedeer — Personalized AI Tutor Prompt — the famous GPT system prompt that turns ChatGPT/Claude into a tutor that adapts to language and level. Use it yourself first to dogfood — read three Ranedeer sessions and you know what a good AI tutor sounds like before you point students at one.
3. Marp — Create Presentation Slides from Markdown — write lecture slides in plain Markdown, export to PDF/HTML/PPTX. Pair with Fabric: extract_wisdom chapter.pdf | tee outline.md then hand-edit into Marp. 30 minutes to a 25-slide deck.
4. notebooklm-py (NotebookLM CLI) — Google NotebookLM grounds answers in your uploaded sources (textbooks, PDFs, lecture recordings). The CLI lets you scriptably create a notebook per unit and generate study guides, briefing docs, or audio overviews students can listen to on the bus.
5. Heyform — Open-Source Conversational Form Builder — self-hosted, GDPR-safe Typeform alternative for formative quizzes, exit tickets, and end-of-unit surveys. Conditional logic lets you branch a quiz based on prior answers — closer to real assessment than 10 multiple-choice in a row.
6. Moodle — Open-Source Learning Management System — the LMS most schools/universities will already accept (or already run). Use it as the gradebook of record and the place to publish the Marp PDFs and Heyform links.
7. Canvas LMS — Open-Source Learning Management System — if your district uses Canvas (most US K-12 do), this is the integration target. Use whichever LMS your school standardised on — don't fork the gradebook into a third tool.
8. Claude Code Agent: Microsoft Study Mode — a Claude Code agent that teaches Microsoft/Azure topics through guided discovery (Socratic questions, not answer-dumping). The underlying pattern — "ask, don't tell" — is the template for the tutor persona you want students using, even outside the MS stack.
9. DeepTutor — AI-Powered Personalized Learning Assistant — multi-agent tutoring grounded in a document you provide. Upload a paper or chapter; students get interactive Q&A, scaffolded hints, and worked examples. The closest open-source thing to "every student gets a TA on demand."
10. Mem0 — Memory Layer for AI Applications — the piece that makes #2/#8/#9 actually personalized. Without Mem0 (or equivalent), your AI tutor re-asks "what's your current level?" every session. With it, the tutor remembers a student struggled with quadratics last Thursday and picks up there on Tuesday.

How they fit together

        Fabric  ─────► extract_wisdom / create_quiz patterns
           │
           ▼
     Marp (slides) ──┐
           +         ├──► published in ──► Moodle / Canvas LMS
     NotebookLM ─────┘                        │
     (study packs)                            │
                                              ▼
                                       Heyform (formative quiz)
                                              │
                                              ▼
                                       gradebook + feedback
                                              │
                          ┌───────────────────┴───────────────────┐
                          ▼                                       ▼
              Mr. Ranedeer  /  Microsoft Study Mode      DeepTutor
                  (system prompts)                   (document-grounded)
                          │                                       │
                          └─────────────► Mem0 ◄──────────────────┘
                                  (persistent per-student memory)

The critical join is LMS + Mem0: the LMS holds the official grade, Mem0 holds the informal context ("Maya finally got long division; don't make her redo it"). Both matter; neither alone is enough.

Tradeoffs you'll hit (AI grading accuracy, FERPA/privacy)

• AI-graded vs human-graded — AI grading for multiple choice and short numeric answers is essentially solved. AI grading for essays is not. Use AI for first-pass rubric scoring + comment generation, then spot-check 20% by hand. Never publish an AI-only essay grade without your review — both because it's unreliable and because in many districts it's policy.
• Moodle vs Canvas — pick the one your institution already runs. Both are fine. Don't fight your IT department on this; the LMS is plumbing, not the lesson.
• Mr. Ranedeer vs DeepTutor — Mr. Ranedeer is a prompt you paste into any LLM (zero install, zero infra, works in a school's ChatGPT-EDU). DeepTutor is a platform that grounds answers in a document you upload (better citations, requires deployment). Start with Ranedeer to learn what you want; graduate to DeepTutor when you need source-grounded answers per unit.
• NotebookLM vs uploading a PDF to ChatGPT — NotebookLM grounds every answer in your sources and cites them inline. ChatGPT will happily invent a quote that sounds right. For student-facing material, grounded > fluent.
• Self-hosting Heyform vs Google Forms — Heyform is open-source and you control the data; Google Forms is one click and your school admin already trusts it. For exit tickets you can use Forms; for anything that touches IEP info or sensitive feedback, self-host.

Common pitfalls (over-trusting AI feedback to students)

• Letting the AI tutor be the only feedback a student sees — AI tutors are great at scaffolding the next step; they are bad at noticing the student stopped trying. Schedule a 5-minute human check-in per student per unit. The AI is a force multiplier, not a substitute.
• Pasting student work into a chat tool without checking the data policy — many districts have explicit rules (FERPA in the US, similar laws elsewhere) that prohibit sending identifiable student data to consumer LLMs. Strip names + IDs before pasting, or use the school's enterprise tier (ChatGPT-EDU, Claude for Education, etc.) which has data-processing agreements.
• Generating a 40-question quiz and not reading it — AI-generated quizzes often have one or two ambiguous items, a duplicate, or a question that requires content you haven't taught yet. Read every generated question before publishing. Twenty minutes saves a class period of "this question doesn't make sense."
• Telling students "the AI graded you a B" — even if it did, frame the feedback as your feedback informed by the tool. Students react very differently to "the computer says" vs "here is what I'm seeing in your work."
• Skipping Mem0 and wondering why personalization is shallow — without persistent memory, every tutor session is a goldfish. The first week feels magical; by week three students are bored of re-introducing themselves. Wire memory in from day one.