The five things only you know

Why the generic prompt gets the generic answer

The model produces the most likely continuation of what it is given. Give it “a starter activity on fractions” and the most likely continuation is the average of every fractions starter on the internet: a pizza, a think-pair-share, forty-five minutes, a projector. Not because the model is lazy, but because you gave it nothing to be specific with.

Everything that would make the activity fit your room lives in your head. The model cannot reach it. So the craft of prompting, for a teacher, is mostly the craft of noticing what you know that it does not, and writing it down.

The five things

There are five, and they cover nearly every case.

1. Who. Year, subject, class size, language background, reading level. “Class 5, 48 students, most read English as a second language” changes the vocabulary, the sentence length and the activity format in one line.

2. What constrains the room. Minutes available, what equipment exists and does not, whether students can move, whether there is a printer. “35 minutes, fixed benches, no printer, one blackboard.” This kills the projector-dependent activity before it is written.

3. What came before. Yesterday’s lesson, last week’s topic, what they already have. “Yesterday they added fractions with the same denominator.” The model now knows where to start and, just as importantly, where not to.

4. The live misconception. The specific wrong idea you saw in their work. “Half of them believe 1/3 is smaller than 1/4 because 3 is smaller than 4.” This is the single most valuable line you can write, because it is the thing no lesson-plan website contains and the thing your lesson actually has to fix.

5. The shape of the output. What you want back, and what you do not. “One five-minute starter. No full lesson plan.” Without this you get the whole thing, forty-five minutes of it, and you spend your evening cutting.

Putting it together

Here is the template as a block you can copy into the prompt file this module ends with:

Class: [year, subject, size, language background]
Room: [minutes, equipment, layout, printing]
Before this: [what they did last lesson]
Misconception: [the specific wrong idea, in the students' words if you have them]
Give me: [exactly what, exactly how many, exactly how long]
Do not: [what you do not want]

Filled in, for an English class:

Class: Form 2, English, 40 students, Kiswahili first language, reading age around 10.
Room: 40 minutes, no projector, I can photocopy one page.
Before this: They wrote a paragraph describing their journey to school. Most paragraphs were a list of events with "and then".
Misconception: They think a good paragraph is a long one. Nobody used a topic sentence.
Give me: Two example paragraphs about the same journey, one that is a list and one with a topic sentence and three supporting details, short enough for a reading age of 10. Then four questions that make students notice the difference.
Do not: Explain what a topic sentence is. I will do that.

That prompt takes ninety seconds to write and produces something you can photocopy. The generic version takes ten seconds and produces something you cannot.

Let it interview you

If you cannot think what to put, reverse the direction. “Before you write anything, ask me five questions about my class and my room that would change what you produce.” The model asks; you answer in a line each; then it writes. This works because the model has read a great deal about what makes lessons fail, and it will ask about the things it needs. It also has a limit: left unsteered it asks generic questions (“what are your learning objectives?”), so add “ask about constraints I might not have mentioned, not about objectives”.

What to leave out

Do not describe your teaching philosophy. Do not explain that you want the activity to be engaging. Do not write “please” and “thank you”: they cost nothing but they add nothing. Every sentence that is not one of the five things dilutes the ones that are, and the memory-limit lesson explained why: material in the middle of a long input is used less than material at the edges. Keep the misconception near the top or the bottom.

The honest limit

Five things make the output fit your room. They do not make it correct. A perfectly targeted starter can still contain a wrong fraction comparison, and the checking habits from the first block apply to every output, however well-prompted. Prompting well reduces the editing. It does not remove the reading.