Hiya Craig - loving these posts and I know exactly how valuable these are. As a physics teacher there are various kinds of diagrams we use in topics about forces, energy, nuclear physics, radioactivity, etc, but the one we need most, and I am still struggling with, is electrical circuits. Knowing whether a student has truly understood the rules of voltage and current by solving circuits of incrementally increasing complexity would be incredibly valuable; and being able to generate particular problems on the fly even more so. Yet no matter which AI I use or the instructions I give it, they cannot consistently generate accurate diagrams. A phantom wire here, a misplaced voltmeter there, a set of currents that don't add up. The day it works is the day my electricity teaching gets better. Still trying...
Couple of tips (apologies if you're already doing this):
1. I would not use image generation for this and instead get the computer to write code to produce the diagram as e.g. a SVG. Prompt it to write a framework/functions first to represent the diagram abstractly in code so it can check the physical constraints, and then a rendering engine to plot it to svg
2. use the codex desktop app with @browser, which will enable the AI to review the diagram after producing it to check whether it's right or wrong
But I suspect that (1) would be enough assuming you're using a frontier model like codex 5.5 or opus 4.7. I'd recommend codex simply because it's a lot cheaper.
Hi Robin, appreciated - I have a Claude Max account using Opus 4.7 so no issues with the specific model! I am already using that approach - I've used Gemini 3.1 Pro to do the same thing, and both hit the same snag - they create the SVGs, but consistently fail to render properly in ways I've already described. I've also tried the image generation approach and it's actually slightly more consistent, although still regularly wrong. I'll keep going - there's going to be a way somewhere.
Hi Craig - loving your enthusiasm as always and I totally agree. Im about to take up a HoD post (again... long story!) And dont want to population our sow with static ppts... we need resources that can adapt to the students in front of you - in real time, using mwb.
Really insightful post, Craig, and a very clear example of why this space is becoming so exciting.
I especially liked the way you framed diagrams not as decorative outputs, but as tools for better mathematical thinking and practice design. The examples around variation, mixed skills and instant worksheets are very powerful, because they show how AI can support teachers beyond simply generating more questions.
This resonates a lot with what we are exploring at Flow IB: how AI can support IB students not only with explanations, but also with diagrams, graphs and visual reasoning as part of the learning process.
The exciting part, in my view, is not just generating more content, but designing better practice, controlled variation, targeted misconceptions, mixed skills and learning sequences around specific goals.
The key challenge is reliability: diagrams must be mathematically accurate, not just visually convincing. But your post makes a very strong case that, if handled well, this could become a major step forward for AI in maths education.
Loving this and the work you are doing. I am currently experimenting with skills in Claude and I am finding this as a great way to have a level of standardisation (i.e deterministic gen AI) balanced with the probabilistic / creative state of LLM models
I'm using Claude AI to generate diagrams, interactive diagrams and animations - that's what I really want for mechanics topics such as projectiles. I find the initial diagrams are a good starting point but there is a lot of tweaking needed such as : "place the bottom edge of the rectangular box on the surface of the inclined plane" (to deal with floating boxes, " move the label for for 'T' to be 10 pixels to the right of the centre of the arrow it corresponds to" - to deal with vector labels on top of arrows ... and so on and so on and so on .... Hopefully "it" is "learning" from my "feedback".
I've chastised Claude a few times with more simple diagrams for MS Forms re-cap quizzes, but haven't delved back into recently.
The point around differentiated (the term *will* come back in a few years) worksheets is excellent and very true and this is where AI genuinely becomes useful. Especially for 'self-selection worksheets'...
I'm sure you have, but have you thought about combining the ideas from the previous post where the AI 'knows' the student to help create even more personalised worksheets?
I'm loving these resources, giving me great practice and starter materials for several different classes. I've been very happy with the coordinate geometry resources including a graph for the distance and midpoint problems. These topics where you want diagrams have been a struggle to find enough resources often, so it is great to have these repeatable questions. I'll have to see what you meant about changing the difficulty, I have to say for formatting it is a blessing to be able to have the worksheet option on every page.
For review starters I often want to combine 3-4 different topics, sometimes across the different "year" levels that you have established on your site as my groups can be stronger in some areas and weaker in others. Am I missing an easy way to do that or should I stick with snipping the questions as I find most useful?
Questions from multiple, unrelated topics is hard, especially when diagrams are involved as it takes so much to get each one right. I have a load of numeracy mixed topic retrieval questions on my site which do some of the work. Check out Plenty in 20 or the Super 8 essential skills
Super 8s have been great, I use it when I want wordier questions than Mathsbot. It can feel a little limiting to only be able to group together "year 8" skills or a specific set of percentages skills, which is why Mathsbot still gets use from me.
But I hear you about the formatting, I can imagine the amount of nightmares you have had to deal with getting to where things are! As I said earlier this amount of flexible resourcing has been amazing, and I've found it easy to supplement with a couple of problem sets made by gemini based off of Dr Austin's stuff. I wouldn't have been doing those sheets without reading your AI posts, so thanks so much. My students definitely have a higher quality and quantity of resources available to them this year compared to the last!
I love the AI resources from you website. Thank you very much for providing them. I agree with Matt. Super 8s and plenty in 20 are very versatile however I feel they missing a few tweaks that would make them amazing for the purpose of low stake quizzing.
1)the ability to select the topic and sub topics of questions like in shedloads
2)combine with other topics like in super 8s
3) produce more or less than 8 questions
4)ensure one of every question selected features in the quiz and have a say in the order?
5)easily generate follow ups such as shadows or shedloads of each selected skill.
For example I may want to test my year 9 ability to: Factorise quadratics by a common number factor, variable factor, double bracket, DOTs, number factor and double bracket, along with some pythagoras, and nth term linear sequences, plus 6 questions for each I can select for addition practice,followed by a restest?
It’s tricky, because each topic (and subtopic) requires careful prompting to ensure suitable constraints, clear valid diagrams, etc. Doing that, essentially for the whole curriculum, means quality will inevitably be compromised.
Hiya Craig - loving these posts and I know exactly how valuable these are. As a physics teacher there are various kinds of diagrams we use in topics about forces, energy, nuclear physics, radioactivity, etc, but the one we need most, and I am still struggling with, is electrical circuits. Knowing whether a student has truly understood the rules of voltage and current by solving circuits of incrementally increasing complexity would be incredibly valuable; and being able to generate particular problems on the fly even more so. Yet no matter which AI I use or the instructions I give it, they cannot consistently generate accurate diagrams. A phantom wire here, a misplaced voltmeter there, a set of currents that don't add up. The day it works is the day my electricity teaching gets better. Still trying...
Couple of tips (apologies if you're already doing this):
1. I would not use image generation for this and instead get the computer to write code to produce the diagram as e.g. a SVG. Prompt it to write a framework/functions first to represent the diagram abstractly in code so it can check the physical constraints, and then a rendering engine to plot it to svg
2. use the codex desktop app with @browser, which will enable the AI to review the diagram after producing it to check whether it's right or wrong
But I suspect that (1) would be enough assuming you're using a frontier model like codex 5.5 or opus 4.7. I'd recommend codex simply because it's a lot cheaper.
Hi Robin, appreciated - I have a Claude Max account using Opus 4.7 so no issues with the specific model! I am already using that approach - I've used Gemini 3.1 Pro to do the same thing, and both hit the same snag - they create the SVGs, but consistently fail to render properly in ways I've already described. I've also tried the image generation approach and it's actually slightly more consistent, although still regularly wrong. I'll keep going - there's going to be a way somewhere.
Hi Craig - loving your enthusiasm as always and I totally agree. Im about to take up a HoD post (again... long story!) And dont want to population our sow with static ppts... we need resources that can adapt to the students in front of you - in real time, using mwb.
Looking forward to using your resources!
Really insightful post, Craig, and a very clear example of why this space is becoming so exciting.
I especially liked the way you framed diagrams not as decorative outputs, but as tools for better mathematical thinking and practice design. The examples around variation, mixed skills and instant worksheets are very powerful, because they show how AI can support teachers beyond simply generating more questions.
This resonates a lot with what we are exploring at Flow IB: how AI can support IB students not only with explanations, but also with diagrams, graphs and visual reasoning as part of the learning process.
The exciting part, in my view, is not just generating more content, but designing better practice, controlled variation, targeted misconceptions, mixed skills and learning sequences around specific goals.
The key challenge is reliability: diagrams must be mathematically accurate, not just visually convincing. But your post makes a very strong case that, if handled well, this could become a major step forward for AI in maths education.
Loving this and the work you are doing. I am currently experimenting with skills in Claude and I am finding this as a great way to have a level of standardisation (i.e deterministic gen AI) balanced with the probabilistic / creative state of LLM models
I'm using Claude AI to generate diagrams, interactive diagrams and animations - that's what I really want for mechanics topics such as projectiles. I find the initial diagrams are a good starting point but there is a lot of tweaking needed such as : "place the bottom edge of the rectangular box on the surface of the inclined plane" (to deal with floating boxes, " move the label for for 'T' to be 10 pixels to the right of the centre of the arrow it corresponds to" - to deal with vector labels on top of arrows ... and so on and so on and so on .... Hopefully "it" is "learning" from my "feedback".
Good job!
I've chastised Claude a few times with more simple diagrams for MS Forms re-cap quizzes, but haven't delved back into recently.
The point around differentiated (the term *will* come back in a few years) worksheets is excellent and very true and this is where AI genuinely becomes useful. Especially for 'self-selection worksheets'...
I'm sure you have, but have you thought about combining the ideas from the previous post where the AI 'knows' the student to help create even more personalised worksheets?
I'm loving these resources, giving me great practice and starter materials for several different classes. I've been very happy with the coordinate geometry resources including a graph for the distance and midpoint problems. These topics where you want diagrams have been a struggle to find enough resources often, so it is great to have these repeatable questions. I'll have to see what you meant about changing the difficulty, I have to say for formatting it is a blessing to be able to have the worksheet option on every page.
For review starters I often want to combine 3-4 different topics, sometimes across the different "year" levels that you have established on your site as my groups can be stronger in some areas and weaker in others. Am I missing an easy way to do that or should I stick with snipping the questions as I find most useful?
Questions from multiple, unrelated topics is hard, especially when diagrams are involved as it takes so much to get each one right. I have a load of numeracy mixed topic retrieval questions on my site which do some of the work. Check out Plenty in 20 or the Super 8 essential skills
Super 8s have been great, I use it when I want wordier questions than Mathsbot. It can feel a little limiting to only be able to group together "year 8" skills or a specific set of percentages skills, which is why Mathsbot still gets use from me.
But I hear you about the formatting, I can imagine the amount of nightmares you have had to deal with getting to where things are! As I said earlier this amount of flexible resourcing has been amazing, and I've found it easy to supplement with a couple of problem sets made by gemini based off of Dr Austin's stuff. I wouldn't have been doing those sheets without reading your AI posts, so thanks so much. My students definitely have a higher quality and quantity of resources available to them this year compared to the last!
Hi Craig
I love the AI resources from you website. Thank you very much for providing them. I agree with Matt. Super 8s and plenty in 20 are very versatile however I feel they missing a few tweaks that would make them amazing for the purpose of low stake quizzing.
1)the ability to select the topic and sub topics of questions like in shedloads
2)combine with other topics like in super 8s
3) produce more or less than 8 questions
4)ensure one of every question selected features in the quiz and have a say in the order?
5)easily generate follow ups such as shadows or shedloads of each selected skill.
For example I may want to test my year 9 ability to: Factorise quadratics by a common number factor, variable factor, double bracket, DOTs, number factor and double bracket, along with some pythagoras, and nth term linear sequences, plus 6 questions for each I can select for addition practice,followed by a restest?
I am curious about you thoughts on this craig?
It’s tricky, because each topic (and subtopic) requires careful prompting to ensure suitable constraints, clear valid diagrams, etc. Doing that, essentially for the whole curriculum, means quality will inevitably be compromised.
I’ll keep thinking