Planning with AI, keeping the judgement · AI for Teachers

The division of labour that works

You decide what the lesson is for. It generates raw material. Those are different jobs and only one of them can be handed over.

You know your Grade 8 class believes plants take food from the soil. You know Fatima finishes in four minutes and needs somewhere to go next. You know the bell goes at 11:40 and the last five minutes are lost to noise. None of that is in the model. All of it determines the lesson.

Put your constraints in the prompt

Weak: "Make me a lesson plan on photosynthesis for Grade 7."

You will get something generic, forty-five minutes long, shaped around a group activity you have no room to run.

Better:

I teach Grade 7 science in a government school. 52 students, fixed benches, 40 minutes, one blackboard, no lab, no projector. Most students read English as a second language. Last lesson they learned that plants make their own food. The misconception I need to break: they believe plants take food from the soil through the roots. Give me three different ways to open this lesson, each under six minutes, each using something students can see or do at their desk. Do not give me a full lesson plan.

The difference is not politeness. The second prompt contains the parts only you know. And note the last line — asking for one component instead of the whole thing keeps you doing the assembly, which is where the judgement lives.

Ask it for what you are worst at

Most teachers ask for a full plan, which is the thing they are already good at producing. That is the wrong request.

What it genuinely helps with, because it draws on far more written teaching than one person can read:

Likely wrong answers. "What will Grade 7 students get wrong here, and what are they thinking when they get it wrong?" You will get eight, and one or two will be errors you had not seen in fifteen years of teaching.
Analogies. Ask for six, reject five. The sixth is often worth the whole exercise.
Questions that expose the misconception. Not questions that test the content — questions where a student holding the wrong idea gives a visibly different answer from one holding the right idea.
A fourth explanation. When you have explained it three ways and Amara still does not have it, another framing at nine in the evening is worth a lot.

What it gets wrong every single time

Timing. It writes 55 minutes of activity for a 40 minute period. Cut a third before you look at anything else.

Checking. It over-produces activities and under-produces the moments where you find out whether anything landed. Add those yourself.

Invented specifics. It will recommend a video by title, a page in a textbook, a named study. Some of them do not exist. Anything with a name, a number or a date gets checked before it reaches students.

Sameness. Left alone it puts think-pair-share into everything, including lessons where nobody can turn around.

The two questions before you use it

1What is each part for? If you cannot say why the third activity is there, delete it.
2If the lesson runs ten minutes short, what do you cut?

Answer both and it is your plan and the machine helped. Fail either and you are holding a script. Scripts collapse the moment a student asks something unexpected, which is about four minutes in.

Before you move on

Two teachers plan the same lesson. Rahel asks for a complete lesson plan and edits it. Nomsa asks for twelve likely student misconceptions and six analogies, then writes the plan herself. Whose lesson is more likely to improve, and why?

Pick the one you would defend. Nobody sees your answer.