The Box and the Module

Why Design Thinking Is the AI Skill Educators Need Most Right Now

One of my most cherished accomplishments in high school was constructing a small wooden box in what was commonly referred to as “shop” in the 1980s. A graduation requirement, the class was held in what today might be considered a maker space but at the time contained a table saw, hammers, and other carpentry equipment rather than 3-D printers. The sole assignment for the semester was to make something - in my case, as someone not especially inclined towards building things, I chose perhaps the simplest project imaginable - a box.

I had to measure the wood, cut the pieces, align the sides, affix the hinges, and sand and polish it until it gleamed. Aside from a research paper I wrote on Andrew Jackson, that box was one of the few assignments I still clearly remember. Unlike the paper, that physical box traveled with me from high school to college, two summers spent living in Japan, law school, and it even made it to my first real apartment, finally getting lost between moves sometime in the early 2000’s. I had that box for almost 20 years, storing change, receipts, and other miscellaneous detritus from daily life, and every time I looked at it I took pride in one of the few things I actually made with my own hands.

I’ve been thinking about that box recently. Not because of any sudden nostalgia for high school, but because I’ve started building things again - this time with AI. And the experience has reminded me that the act of making something, of designing and constructing something that didn’t exist before you willed it into being, engages a part of your brain that merely asking for information does not.

From Prompting to Building

In a recent post, I argued that educators should approach the new capabilities of agentic AI with imagination as well as caution. That piece was about what these tools can do. This one is about what kind of thinking they actually require - and why that thinking might reassure educators rather than alarm them. What I didn’t fully appreciate at the time was how much that thinking would resemble the work teachers already do every day.

The shift I’ve felt over the past several months isn’t entirely new, but one of degree. Agentic tools - AI that can build entire applications and websites rather than just generate text - have been available since at least the middle of 2025, but since the new year they've become significantly easier for non-coders to use. What’s changed for me is the focus: from generating content to designing experiences.

When I first started using AI seriously, I was doing what most teachers do - asking questions, iterating on responses, but primarily generating text - converting ideas into polished plans and getting feedback and analysis on my own work. AI remains useful for all of those tasks. But agentic AI introduces something fundamentally different. I can now craft an interactive learning experience - deciding what it should do, how it should be organized, who would use it, and what problem it needs to solve. That’s a much different - and more involved - cognitive activity.

Design Thinking

About a decade ago, our school brought in a consultant from IDEO - the design firm that essentially pioneered what’s now called “design thinking” - for a professional development workshop. I remember finding the session interesting: its emphasis on starting with user needs, prototyping repeatedly, and treating failure as information to act on rather than giving up in defeat. Like most PD workshops, it was engaging in the moment and then I went back to doing roughly the same things I’d always done. Every teacher knows this feeling.

But the notion of “design thinking” stuck with me and now seems more relevant than ever. A core insight - that how something is built and presented to its users matters as much as what it contains - turns out to be a helpful way to approach AI now that teachers have the tools at their fingertips that allow them to act on it.¹

Design thinking is not an alien concept for teachers. Every unit you’ve ever planned, every set of learning objectives you’ve ever sequenced, every decision you’ve ever made about what to include and what to cut are design decisions. We’ve been doing this work our entire careers. But for most of us, unless you happen to have a degree in computer science, using technology in the classroom has been a mostly passive experience. The products and platforms we used were designed by someone else.

We worked inside whatever template Blackbaud or Canvas or Schoology provided and made the best of it. During the pandemic, many of us became more adept at leveraging our LMS through embedded videos, links to platforms like EdPuzzle or Desmos, all in service of organizing our materials online for student access. But we were constrained by the limitations of our school’s chosen product.

The difference now is that teachers can create their own bespoke LMS, limited only by their imagination. That is new and genuinely exciting.

What Design Might Look Like

For years, I’ve run a Constitutional Convention simulation in my U.S. history class. It’s always been a favorite - students are assigned state delegates in attendance at the 1787 Philadelphia gathering, debate the structure of the new government, and vote on the major compromises that shaped the Constitution. I've run this activity for more than 20 years and refined it considerably.

The materials I provide are detailed but sprawling. Students each have a delegate position sheet, primary and secondary source material, and a variety of other excerpts from digital and analog resources. Managing the various documents can be overwhelming, even when students get them ahead of time. This year’s convention went smoothly, but as any teacher knows, there is always room for improvement.

Curious about what was possible, and off the success of several earlier builds using Claude Cowork, I set out to create a single, self-contained online module that would house everything for next year.

Using a dedicated Claude Project, I worked through how to organize and display the material. I settled on a series of easily navigable tabs, each containing exactly what students needed at that moment. That decision required knowing how the activity actually flows in a classroom, which information students reach for and when, and where the bottlenecks have historically been.

I had to decide what belonged inside each section. The Delegate tab needed more than biographical information - it needed each delegate’s positions on every voting issue, organized so students could quickly find where their representative stood. But I also realized that understanding the Convention requires seeing how delegates diverged on the same questions - disagreements about government structure, determining representation, and the limits of executive power. During the design conversation, I described that specific pedagogical need, and Claude suggested a comparison feature where students could pull up two or three delegates side by side.

You can explore the feature and the entire site here.

The comparison feature only existed because I articulated a teaching problem. That pattern repeated at every stage. Whether primary source documents should live inside the module or be linked out (when students click external links, the risk of distraction soars - if the source isn’t right there, it doesn’t get read). Whether my voting spreadsheet could be replaced with a live tracker built into the module itself (it could, and could be downloaded at the conclusion of each day’s session).

None of these decisions required coding knowledge. All of them required teaching knowledge. AI handled the execution while I managed the design. The result is a functioning interactive web application, and early feedback on the modules I’ve used this year suggests it will be an upgrade.

Two Kinds of AI Use

The design dimension of agentic AI changes the conversation in ways that haven't been fully appreciated. To see why, you have to distinguish between two fundamentally different kinds of AI use that are often treated as if they’re the same thing.

The way most students approach AI is one-shot or, at best, slightly iterative. They type a question to get an answer, ask for an essay draft, or request feedback on a problem. This is the cognitive offloading conversation we’ve been having for more than three years. The emerging research confirms what most teachers already suspect - when students use AI as a question-and-answer machine, the offloading is real and any learning gains, if they happen at all, are ephemeral.² Students generate a product without doing the intellectual work that product is supposed to represent.

But designing with agentic AI is a fundamentally different activity. When a user sets out to build something - an interactive module, a web application, or some other tool to be digitally deployed - the thinking has to be front loaded or the result is useless. You can’t open an agentic tool and type “make me a Constitutional Convention website” and get anything worth sharing with students. You might get a generic approximation - perhaps not awful, but mostly another version of AI slop.

The tool needs design specifications that reflect real expertise: what the activity is for, how information should be organized, what the user needs at each stage, and what problem the whole thing is solving. Without that, the AI produces something that may look impressive but serves no one.

My experience over the past several months suggests that the cognitive offloading concern inverts when you shift from content generation to design. Design-based AI work is cognitively demanding in exactly the ways educators value - strategic planning, audience awareness, thoughtful information integration, and the ability to evaluate whether something actually works the way the user intended. The process of designing with these tools isn’t a shortcut around thinking. It’s a more complex kind of thinking than what most traditional assignments ask students to do.

Why Design Rewards Careful Thinking

There’s a practical dimension to this shift that I didn’t fully appreciate until I’d been building for a while.

When you’re chatting with an AI - asking questions, generating text, iterating on a response - the iteration is immediate and forgiving. If you don’t like the output, you try again. You can afford to think loosely because the cost of experimentation is essentially zero. This is what most users have been doing for the past several years.

Design-based work doesn't operate that way. When you're building something with multiple interconnected components, specifications need to be thought through carefully or the features don't fit. This includes asking whether what you're considering is worthy of an online module in the first place. Good questions matter as much as good answers, and careful follow-up matters even more. All of this needs to happen before you ask the model to start building.

You can iterate your way to a good result, but every cycle of build, review, and rebuild costs time and adds complexity. Wobbly thinking at the start means rework at every subsequent stage.

Anyone who has ever renovated a kitchen understands this principle intuitively. You can tear out a wall and start over. But nobody who has done an expensive home improvement project thinks that’s a good plan. You hire an architect to inspect the space and make those hard decisions carefully before you ever start demolition. The economics of construction - whether physical or digital - reward careful upfront thinking and punish sloppy work.

A Note on Students

I’ve been mostly focused on the teacher side of AI use for a reason. The overwhelming majority of students are not using AI this way - not because they aren’t capable, but because they either don’t know agentic tools exist or have no incentive to use them. We don’t typically assign students to build interactive modules, so the motivation to use agentic AI for cheating in this context just isn’t there.³

I'm actually less worried about student agentic AI use than I am about basic chatbot use. A student who could design and build a working module - who could think through the architecture, map the content to a user’s needs, and make deliberate decisions about what to include and how to organize it - would have demonstrated real intellectual work. If a student showed up in my class having built something like that, my reaction would not be to worry about their academic integrity. It would be to ask them to show me how they did it.

This could change. We may reach a point where a two-sentence prompt delivers something spectacular and the design work gets automated along with everything else. But right now, the cognitive demands of designing with agentic AI are creative and significant enough that I think it offers more of an opportunity than a threat.

The Box and the Module

Building an online teaching module with AI is clearly not the same as building a wooden box with your hands. There’s something about physical construction - the weight of the material and the permanence of the object - that represents tangible evidence you made a thing that exists in the world. Digital creation doesn't replicate it.

But the pride I felt in making that box came from the same place as what I feel when I look at the Convention module: I made something I didn’t know I could make. The box taught me that I could measure, cut, align, and finish a physical object despite having no particular aptitude for building things. The module taught me that I could architect, organize, sequence, and deploy a digital one.

AI is rapidly becoming a design medium. It extends what a thoughtful practitioner can do, in the same way that a table saw extended what a clumsy teenager could build in a shop class in 1985. The quality of what comes out still depends entirely on the quality of the thinking that goes in.

Will teachers be willing to think of themselves as designers? The skills required to do that well are the skills most good teachers have spent their careers developing. The only thing that’s new is what those skills can now produce.

Going Forward in 2026

2025 was supposed to be the year of agentic AI. In 2026, it’s actually arriving. The tools are doing many of the things their proponents suggested they would - AI is being integrated into more workflows, gaining the ability to handle longer and more complex tasks, and becoming increasingly accessible to people without technical backgrounds. That’s a significant development, and I’m trying to explain what it actually feels like from inside a classroom - not as a prediction about where things are heading, but as an account of what’s possible right now.

The current conversation about AI in education is still largely split between two worlds. In one, AI companies demo features to administrators and pitch adoption at scale. In the other, schools struggle with the logistics - procurement decisions, acceptable use policies, and faculty training that never quite keeps up. And that’s for schools actually engaging at all. What's been harder to find are concrete use cases - a teacher sitting down with these tools and building something for a specific group of students.

The gap between what I’m describing and where most teachers are is significant. The pedagogical design instinct transfers - most experienced teachers already think this way about their materials and activities. But the technical fluency required to translate that instinct into a built product is real, and bridging it requires sustained investment in time and development that most schools are not yet making. I’ve had the curiosity and the circumstances to develop that fluency. Most of my colleagues haven't. But if the trajectory of the models over the past six months is any indication, that gap will start closing fast.

Yes, the concerns about cheating, privacy, security, and whether AI will ultimately improve learning are real and must continue to be debated. I’ve written about many of them. But I’ve also been paying attention to what these tools actually require when you use them seriously, and what I've found is that using AI effectively is more demanding and more intellectually rigorous than most of the conversation assumes. If that's true, the question isn't whether we need better chatbots. It's whether we need better assignments. I think we will.

The window to shape how these tools get used in education remains open, and the people willing to experiment - not caught up in predicting the future or surrendering to a predetermined narrative, but actually building things for their students and sharing what works and what doesn't - are the ones who will figure it out first.

If design thinking was ever going to find its way into how teachers use AI, this is when it starts.

---

Connect With Me

Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogically questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.

1. IDEO now has courses specifically for how to use AI through the lens of Design Thinking.

2. A recent Stanford review of over 800 studies on AI in K-12 education found that only 20 produced strong causal evidence. The results suggested that AI often acts as a cognitive crutch that harms independent reasoning. Notably, the studies reviewed focused overwhelmingly on chatbot-style interactions, not the design-based use described in this piece. For an excellent practitioner analysis of the Stanford findings, see Wess Trabelsi's recent overview.

3. I realize there has been some recent coverage of agentic browsers and other AI tools that purport to complete online coursework but I’m skeptical these are widely used and, as a K-12 teacher who sees my students in person, not especially concerned high school students will use AI this way.