You ask for a lesson plan on fractions and get something that reads like it was assembled from every lesson plan ever written: technically correct, faintly familiar, and not usable without twenty minutes of rewriting. The instinct is that you need a better tool. Usually you need to give the tool more to work with.
Why the default output is average
A language model predicts likely continuations of text. Asked a broad question, the most likely continuation is the most typical answer, the average of everything it has seen. Average is exactly what "generic" means.
This is not a flaw to be fixed with a cleverer tool. It is what the system does when you haven't told it enough to narrow the field. Every piece of context you add rules out a swathe of plausible-but-wrong answers, and the output moves from the middle of the distribution toward the specific thing you actually wanted.
A generic answer is usually a correct response to a generic question. The model gave you the average because you described the average.
The four things it almost always needs
You do not need elaborate prompt formulas. In practice, most of the improvement comes from supplying four things that teachers know and models cannot guess.
1. Who the students are
More than "Year 8". What they already know, what they struggle with, what they did last lesson. "Year 8, mixed ability, they have met equivalent fractions but not yet operations with them, and about a third are still shaky on times tables" produces something entirely different from "Year 8 maths".
2. The constraint you're working inside
Fifty minutes. No printing. One set of laptops shared between two classes. Last period Friday. Constraints are the most useful information you can give, because they eliminate the largest number of unsuitable answers, and a model with no constraints will cheerfully design a ninety-minute lesson requiring a class set of iPads.
3. What "good" looks like to you
Models have no idea what you value. If you want students talking to each other, say so. If you want a silent starter, say that. If you have a departmental structure you always follow, describe it. Otherwise you get the model's implicit house style, which is generally American, activity-dense, and heavy on group work.
4. An example of what you mean
This is the highest-leverage one and the most often skipped. Paste in a lesson plan, worksheet or rubric you already like and say "match this format and level of detail". One example does more than several paragraphs of description, because it carries all the things about your preferences that you wouldn't think to articulate.
Before and after
Write a lesson plan on persuasive writing for Year 8 English.
Year 8 English, 50 minutes, mixed ability class of 26. They have already done a unit on language devices and can identify ethos/pathos/logos, but their own writing still defaults to listing arguments rather than building one.
Design one lesson focused specifically on structuring a persuasive argument, not on devices, which they've covered. I want at least fifteen minutes where students are writing rather than listening. No printing available; everything must work on the board or on their laptops.
Follow the format of the lesson plan below, including the timing column and the "anticipated misconception" row. [paste your usual plan]
The second prompt is longer, but almost all of it is information you already have in your head. The thirty seconds spent writing it down is what separates a plan you can teach from a plan you have to rewrite.
Iterating instead of restarting
When output isn't right, the common instinct is to rewrite the prompt and start over. It is usually faster to correct what you have.
- Be specific about what's wrong. "The starter is too passive. Replace it with something where every student writes an answer in the first two minutes" beats "make it more engaging".
- Ask it to cut, not add. Models pad by default. "Reduce this to what fits in 50 minutes, and tell me what you removed" is a useful instruction.
- Ask for the reasoning. "Why did you sequence it that way?" occasionally surfaces a good rationale, and more often reveals that there wasn't one, which tells you it's yours to decide.
- Ask for options. "Give me three different ways to open this lesson" is more useful than one confident suggestion, because choosing between three is a task you can do quickly and well.
Write one paragraph describing each class you teach: prior knowledge, common gaps, what works with them, the constraints of the room. Keep it in a note. Paste it at the top of any prompt about that class. You write it once and it improves every output for the rest of the year.
What better prompting will not fix
Context makes output more specific. It does not make it correct. A well-prompted model still fabricates dates, misstates a syllabus outcome, and produces confident nonsense in areas where its training was thin. Specificity raises the quality of the draft; it does not remove the need to read it.
It is also worth being honest about where the ceiling is. The best-prompted output is a competent draft that reflects the average of good practice, shaped toward your context. It is not going to contain the insight that comes from having taught this topic to this class for six years. That part is still yours, and the reason to get better at prompting is to spend less of your evening on the parts that aren't.