AI for Teachers: I Prep a Complete Lesson Sequence with Gemini, Here Is My Exact Workflow

Every Sunday evening, the same scene: the planning journal open, the curriculum to cover, and that recurring question. How do you prepare a solid lesson sequence without spending the whole night on it? The landing pages of AI tools promise “lessons in seconds”. After more than fifteen years in the classroom and daily use of Gemini in CM2, I can tell you one thing: those promises are marketing, not pedagogy.

This article shows the real workflow, the one I use in a class of 25 students, half of them French learners, at the Lycee Francais de Kuala Lumpur. No carefully cherry-picked screenshot, no anonymous testimonial. Copy-ready prompts, the steps before, during and after the sequence, and above all what AI systematically gets wrong.

Key Takeaways

• AI for teachers does not replace preparation: it removes the grunt work (wording, variation, reformatting) and frees up time for pedagogical decisions.
• The workflow has three phases: upfront, Gemini structures the lesson plan and the differentiation; during class, it produces materials; afterwards, it helps with grading and the review.
• The prompts given here are copy-ready as-is and adaptable to any level, from cycle 3 to high school.
• Two things NEVER go through AI: students’ personal data and the final decision about what happens in class.

Bottom line, in the first person: Gemini does not prepare my lessons for me. It prepares me to prepare them. The difference is fundamental, and it explains why I spend less total time on preparation while producing richer documents than before.

Why “1-click lesson generator” tools send me running

The first results on this topic are commercial pages. One promises a lesson “in seconds”, another stacks eight tools to cover a week. None shows a teacher at work, with real constraints: a grade-level meeting on Tuesday, heterogeneous students, an inspector who will visit the classroom in three weeks.

The problem with these tools comes down to one sentence: they produce a document, not a sequence. A lesson sequence is a progression thought through over time, with objectives tied to the curriculum, diagnostic and summative assessments, moments of consolidation. A generator produces a stack of nicely formatted worksheets, often wrong on the concepts, always generic about the students.

I tested several of these platforms last year. Constant result: I spent more time fixing content errors than writing things myself. A worksheet on fractions where the number line was badly built, an informational text on the French Revolution with inconsistent dates. The appealing formatting masked fragile content.

Those landing pages target institutions, not AI for working teachers. My rule is therefore simple: I do not use a tool that hides the reasoning from me. Gemini used in conversation, with real back-and-forth, lets me keep control of every decision. It is the nuance I have defended since I analyzed what delegating to an AI agent really changes: delegate the execution, never the responsibility.

My workflow BEFORE the sequence: the plan with Gemini

Step 1: I provide the official framework, never the other way around

The first classic mistake is asking “prepare a lesson sequence on fractions for me”. The model will improvise a framework. But my framework is the curriculum, my yearly progressions, and the real level of my students. So I start by giving the context.

Here is my opening prompt, copy-ready as-is:

You are a teaching advisor specialized in cycle 3 mathematics. Here is the part of the official curriculum about fractions in CM2: [paste the excerpt]. Here is my yearly progression: [paste]. Here is my class profile: 25 students, about half working on French language support, 5 students comfortable in math, 2 in great difficulty. Do not write anything yet. First ask me 5 questions to better understand my constraints.

This prompt changes everything. By asking questions before producing, Gemini becomes an interlocutor instead of a worksheet machine.

Step 2: the lesson plan, session by session

Once the answers are in, I ask for the skeleton. Always in the same order: objective of each session, activity type, materials, assessment format. The plan fits on one page. I reread it, I delete, I reorder. This is where quality is decided: a lesson plan revised ten times by me beats a brilliant plan I have not digested.

I always specify a format constraint: “One table per session, 5 rows maximum, no written-out paragraphs.” Without that constraint, the model writes descriptive filler that pollutes the review. With it, I get a view I can scan in thirty seconds.

Step 3: differentiation, where AI is actually useful

Differentiation is the most time-consuming preparation task. Three levels of worksheets, visual supports for students struggling in French, a challenge for those who finish early. It is variation work, not creation work. In other words, exactly what a language model does well.

My go-to prompt: “Starting from exercise 3 of session 2, produce three versions: a support version with simpler numbers and a worked example, a standard version, and an extension version that introduces percentages. Keep the same instruction worded identically in all three versions.”

As-is, these prompts turn Gemini into a genuine AI assistant for teachers: it adapts, it rephrases, it does not decide. The gain is real and qualitative: I produce three levels of worksheets where, before, I produced two for lack of time. My advanced students are no longer bored, and my struggling students no longer face an exercise they cannot read. For the French-learner half-group, I also ask for shortened instructions with controlled vocabulary.

Step 4: the in-class materials

Projector on, I generate the number-line plots in LaTeX or as a precise description, the contextualized problem statements (I localize them to Kuala Lumpur, the Petronas Towers replace the Eiffel Tower, and my students relate to them better), and the self-assessment grids. Every material is proofread before printing. The proofread takes ten minutes; it is non-negotiable.

My workflow DURING the sequence: AI barely enters the classroom

A crucial point that marketing demos gloss over: during class, AI is absent. I lead the class, I observe, I adjust. I only use Gemini in two specific cases, between two sessions or at the end of the day.

First case: the pedagogical surprise. Session 3 revealed a massive misunderstanding about adding fractions. That same evening, I describe the situation and ask for three targeted remediation exercises for the next day. Five minutes of work, one relevant exercise for the next morning.

Second case: group regulation. When a conflict or a class dynamic raises a question, I describe the situation cold, with no identifying data, and ask for leads. The rephrasing exercise helps me decide on my own.

What I never do: project generated answers without having read them, or let a student chat with a chatbot in class without a framework. A minor’s data never enters a request. No first name, no identifiable work sample. It is a red line, consistent with the usage recommendations in Gemini for education published by Google: the tool is aimed at the professional, not at students directly.

My workflow AFTER: grading, review, anticipation

Grading, partially assisted

For grading, Gemini helps me on two limited tasks. First, I build the detailed answer keys and the grading scales from the problem statements, which is fast and secures my consistency across the three differentiation levels. Then, for assessments, I write standard comments by error profile: I describe three observed mistakes, anonymized and rephrased, and ask for three constructive feedback phrasings.

I do not do automatic grading of student papers. I read everything. A CM2 student’s paper is a window into their reasoning, and reading it is my job. What AI gives me is speed on the administrative side of feedback, not the judgment.

The end-of-sequence review

At the end of a sequence, I feed a personal note: what worked, the sticking points, the students under strain. Gemini helps me turn those raw notes into a structured review for the grade-level meeting or my own progression next year (N+1). When I compare Gemini and ChatGPT for this kind of work, long-document rephrasing quality is what makes the difference for this review.

Summary table: what I delegate, what I keep

This table sums up my doctrine: the more a task is variation, the safer the delegation. The more it touches pedagogical judgment, the more I take back control. AI for teachers is a usage problem before it is a tool problem.

The three things AI systematically gets wrong

After two years of practice, three blind spots keep coming back.

First, the real level of the students. A language model knows the theoretical CM2 of the curriculum, not my 25 students of this year. It almost always overestimates what they can read on their own. My generated instructions are systematically cut in half, and the vocabulary checked against the classroom tools.

Next, the temporal coherence of the sequence. Gemini tends to stack rich but unrealistic activities in terms of time. Preparing a sequence also means pacing real classroom time, which a model ignores. A 55-minute session with three hands-on phases, a written trace and a group debriefing does not fit. I am the one who cuts, and that is perfectly normal: only the teacher knows what “debriefing as a group” means with their class.

Finally, the sensitivity of the local context. The model does not know that some of my students discover metropolitan France through its textbooks, nor that the Malaysian school calendar shifts my weeks. I provide this information, and AI becomes relevant. Without it, it produces generic output. Preparing a lesson sequence is situated work, and AI knows nothing about your situation until you tell it.

My safeguards: three non-negotiable rules

Rule 1: never a student’s identifying data in a prompt. I talk about “student C” or anonymized profiles. It is a discipline, not an option. It protects my students and it protects me professionally.

Rule 2: every generated resource is proofread before class. No exceptions. A badly drawn figure, a wrong date, an ambiguous instruction, and it is my teaching credibility that takes the hit. The ten-minute proofread is the real cost of delegation.

Rule 3: AI proposes, I dispose. Every suggestion is a hypothesis to validate, not an instruction. If I cannot justify a session in front of a colleague or an inspector, it does not enter the planning journal, whatever its apparent quality.

These three rules hold well beyond school: my work on this site follows exactly the same principles.

Frequently asked questions about AI for teachers

Is Gemini free for teachers?

There is a free version that is enough to get started, and education-specific offers depending on the school. I work with the general public version and the core features cover most of the workflow described here: long-form writing, multi-turn conversation, table generation.

Can AI grade student work?

I advise against fully automatic grading. That said, AI helps build answer keys, grading scales and feedback phrasings. The judgment on a student’s paper remains a human teaching act, especially in primary school where the paper reveals reasoning under construction.

How do you write a good lesson-sequence prompt?

Three ingredients: the official framework pasted in full, the real profile of the class, and a strict format instruction. Add the instruction “ask me questions before producing”: it turns a generic answer into a useful exchange. A good prompt runs ten lines, not one sentence.

Does AI replace school textbooks?

No, and that is not the right debate. The textbook brings a validated progression, calibrated exercises and a consistency that I keep. AI works around the textbook: adapting, localizing the context, creating the supplementary materials. The two coexist in my classroom without tension.

Should primary school students be taught about AI?

In CM2, I talk about AI occasionally and with supervision, to lay the foundations of critical thinking: a machine can be wrong with total confidence. But the tool remains an instrument of the teacher at that age. Real training comes later, with progressive uses in middle school and high school.

Final word

Preparing a lesson sequence with Gemini looks nothing like what the ads promise. No lesson in one second, no magic. A disciplined assistant, used before and after class, that removes the variation work and gives me back time for what matters: observing my students, adjusting, explaining.

If you are a teacher and you are hesitating, start small. One single sequence, one single planning prompt, the full proofread. You will quickly see where delegation serves you and where it hurts you.