The second message matters more than the first

Where teachers spend their effort, and where it pays

Most people meeting these tools put all their effort into the first prompt. They write it, read the output, feel disappointed, and rewrite the prompt from scratch. Three rounds of that and they conclude the tool is not much use.

The effort is in the wrong place. The first output is almost always a draft. The fastest route from draft to usable is not a better first prompt; it is a pointed second message. A model handed a specific critique of its own output improves markedly, because the critique tells it exactly which part of the space to move to. A rewritten first prompt starts the guess again from nothing.

What a useful second message looks like

Three parts: what is wrong, where, and why.

Weak: “Make it better.”

Weak: “These are too easy.”

Useful: “Questions 1 to 6 can all be answered by copying a phrase from the passage. I want questions where the answer is not in the text and the student has to infer it. Question 4 is close: do the rest like question 4.”

The last version names the fault, points at the location, gives the reason, and hands over an example of the target from the model’s own output. That last move is the strongest available: you are showing it one of its own, which the previous lesson explained is the densest steer there is.

Why “better” fails mechanically

“Better” has no direction. The model resolves it to the most common meaning of better in the text it learned from, which for generated content is nearly always “more”: more questions, longer explanations, extra sections. You asked for a change of kind and got a change of quantity. This is not the model misunderstanding. It is the model correctly guessing what an unspecified “better” usually means.

Every vague critique has a default the model falls into. “More engaging” produces exclamation marks and a game. “Simpler” produces shorter sentences with the same concepts. “More rigorous” produces longer sentences with the same concepts. If you find the same disappointing pattern each time, the pattern is the default, and the fix is to say what you actually mean.

One thing at a time

A critique that lists six problems gets six half-fixes. The model addresses each, briefly, and the output becomes a compromise. Fix the biggest problem first, look at the result, then fix the next. Three focused rounds beat one comprehensive complaint, and they take less time in total because each output is easier to read.

When to stop and start again

Iteration has a limit, and the memory-limit lesson explained it. After eight or ten rounds the conversation is long, the early constraints are drifting out of the window, and the model is carrying the residue of every abandoned version. Signs that you have reached it: the output starts reintroducing things you removed three messages ago; the model apologises and produces the same thing again; the tone shifts.

At that point, take the best version, open a new chat, paste the best version as your example, and write a clean prompt around it. You lose nothing (the good output travels with you) and you get a model with a clear window.

The critique is teaching you too

There is a side benefit that experienced teachers notice quickly. Writing “these are all recall, I want inference” forces you to name what you wanted, in words. Teachers who spend a term critiquing generated questions become sharper at seeing the difference between recall and inference in their own questions, because they have had to articulate it two hundred times. The tool did not teach them that. The act of correcting it did.

A pattern to copy

For any generated resource, this three-message shape covers most cases:

1. The five-things prompt.
2. “Questions [numbers] have problem [X] because [reason]. Do them like question [Y].”
3. “Now check every question against the passage and tell me any whose answer is not actually in it, or is ambiguous.”

The third message is a self-check. It is not reliable (the model checking its own work misses things a person would catch) but it catches some, cheaply, before you read. Then you read.

The honest limit

A model steered by critique can circle. Ask it to make questions harder and then more accessible and then harder again and you will get an oscillation rather than a synthesis, because each message pulls toward one pole. When two goals pull against each other (rigour and accessibility, brevity and completeness) state both in the same message and let it find the balance in one go. Sequential pulls do not converge.