Using AI to plan lessons without handing over your teaching
The useful question about AI and lesson planning is not whether it works. It is which parts of planning it can take, and which parts you should keep.
After a few hundred generated lessons, the split turns out to be fairly clear.
What AI is genuinely good at
Structure and scaffolding. Given a topic, a level and a duration, a model will produce a sensible progression — warm-up, presentation, practice, production. This is the boring, repetitive part of planning, and it is the part AI removes almost entirely.
Volume. Twelve gap-fill sentences using the present perfect. Twenty vocabulary pairs. Six comprehension questions at B1. Generating variations is mechanical work, and a model does it in seconds without getting bored on the eleventh item.
Getting past a blank page. A mediocre first draft you can edit is worth more than a perfect lesson you have not started. This is the single biggest time saving, and it is mostly psychological.
Differentiation. "Same topic, one level down" is a genuinely tedious task done by hand and a trivial one for a model.
What it gets wrong
Your specific student. A model does not know that Marco freezes when asked to speak first, or that this particular class has already done three lessons on transport and is sick of it. Everything about pacing and sensitivity to the individual is yours.
Level calibration. Models drift upward. Ask for A2 and you will often get something closer to B1 — vocabulary slightly too broad, sentences slightly too long. Assume you will need to simplify.
Factual detail in specialist subjects. For history dates, scientific mechanisms or anything where being wrong matters, verify. A confidently phrased error is worse than no material, because it reads as authoritative to a student.
Cultural fit. Generated examples skew towards a generic Anglophone context. If you teach in a different setting, expect to swap examples.
Knowing when to stop. AI will happily generate a fourth exercise on the same point. Judging that the student has had enough is a teaching decision.
Writing a brief that produces something usable
The quality of what comes back is almost entirely determined by the brief. Most disappointing output is a vague-request problem, not a model problem.
A weak brief:
Make a lesson about the past tense.
A brief that produces something you can actually teach:
Subject: English. Topic: past simple for finished actions. Level: A2. Duration: 45 minutes. Focus: irregular verbs the student keeps getting wrong. The student should practise producing full sentences aloud, not just filling gaps.
Four things make the difference:
- Level and duration, always. Without them the model guesses, and it guesses long.
- A narrow topic. "Past tense" is a syllabus. "Past simple for finished actions" is a lesson.
- What the student should practise, phrased as an activity rather than a subject. This is the highest-leverage sentence in the whole brief.
- One constraint from the real world — what they struggle with, what you have already covered, what to avoid.
Nothing in that brief identifies the student. It should not: there is no reason to send a name, a grade or a submission to a model in order to get a lesson plan back. In Cleo AI the request contains only the subject, topic, level, duration and focus — the material is generated from the teaching parameters, not from student records.
The edit pass that matters
Treat generated material as a draft from a competent but unfamiliar colleague. Three checks, in order:
Cut by a third. Generated lessons are consistently too long for their stated duration. Removing the weakest exercise usually improves the lesson.
Check the level down, not up. Read the hardest item and ask whether your actual student could do it unaided. If not, simplify — do not add a supporting explanation, which lengthens the lesson.
Replace one example with something real. A single example drawn from your student's life or your local context changes how the whole lesson lands. This takes a minute and is the highest-value edit you will make.
Where this leaves the teacher
The parts of teaching that AI cannot do are, conveniently, the parts that were always the actual job: reading the room, deciding when to move on, knowing that a student's third wrong answer means something different from their first.
What it removes is the assembly work — the typing out of twelve sentences, the formatting, the rebuilding of a structure you have built ninety times before. That work never taught anyone anything, and there is no craft lost in automating it.
The teachers who get the most out of this are not the ones who accept the output. They are the ones who write a sharp brief, cut hard, and put their own example in.
A reasonable place to start
Take a lesson you have already taught successfully. Write a brief for it and compare what comes back with what you built. You will see immediately where the model is strong and where your judgement was doing the work.
That comparison is worth more than any amount of advice about prompting, because it calibrates you on one specific thing: how much of your planning was craft, and how much was typing.