--- slug: not-another-webapp title: "Not Another Webapp" date: 2026-08-30 dek: "We built an AI product for teachers, went up against hundreds of engineers and billions of dollars building the same interface, and shut it down. The bet we're making now isn't another app." --- As part of our go-to-market for an ed-tech product, we ran professional development sessions with schools on generative AI. We'd demonstrate ChatGPT, Gemini, Claude, and, of course, our own product. It was a useful way to show what was possible with generative AI, while also testing our thesis about what a tool built specifically for educators could do differently. We had two bets. The first was that we could build a simple, intuitive interface for rich resource creation. The second was that we could build a connected, project-based knowledge system that gave language models the context they needed to produce genuinely useful work for educators. The first bet worked. Our product could generate teaching resources, but not through the familiar collection of buttons, dropdowns and forms. It was native to a chat interface: a conversation produced a rich-text artifact that you could edit, argue with, and rework. We had a strong intuition about what happens when you let an agent work directly inside a document. We were right. We were just early. That interaction is now everywhere. The interface we thought might differentiate us became a commodity, and the math stopped working. We were a two-person startup competing with hundreds of engineers and billions of dollars of investment to build essentially the same interface. Pricing was another problem. Education doesn't really run on ROI. A teacher isn't usually asking whether an AI tool will generate a measurable return on investment. They're asking whether it will make Monday morning a little easier. Schools have procurement processes, budgets and competing priorities, but the person actually using the product is often making a much simpler calculation: *is this useful to me?* We were also trying to price a product built on top of a foundation model whose capabilities and economics were changing underneath us. And we weren't only competing with other startups. We were competing with state departments building their own AI products, backed by public money. So we migrated our users off the platform and shut it down. ## The missing ingredient is context Ask a general-purpose model to write a lesson plan aligned to a particular curriculum and it can produce something that looks convincing. It might even cite curriculum outcomes that sound exactly right. We have even seen our prior competitors publish marketing material with incorrect references. That's not necessarily a model failure. It's a context failure. The model can be excellent at generating the lesson plan. What it doesn't necessarily have is access to the particular curriculum, scope and sequence, school policies, assessment frameworks, local terminology, existing resources, or the accumulated knowledge of the teacher sitting in front of it. The problem isn't that AI can't generate the answer. It's that the useful answer is usually dependent on knowledge that isn't in the model. This isn't unique to education. Every industry has its own context. But education makes the problem particularly obvious because so much of that context is distributed across institutions, jurisdictions, curriculum documents, teachers and classrooms. And unlike many enterprise settings, the people who hold that knowledge aren't necessarily the people with the budget to build software around it. At the same time, the potential value of AI in education is enormous. So we've been asking a different question: **what if the answer isn't another application?** ## From applications to primitives Since shutting down the product, we've been iterating on the second bet: building the beginnings of AI infrastructure for public context, starting with education. Not another destination where a teacher goes to "do AI." Instead, we're interested in the primitives that make AI work naturally with the context educators already have. That has meant experimenting with agent skills, connectors and plugins — small pieces of infrastructure that can give an agent access to specific capabilities, workflows and sources of knowledge. The skills and plugins side is newer, and I don't pretend we've figured it out. I use skills constantly in my own engineering work, and I've been experimenting inside the school where I work, building small automations around the things educators actually need. What I don't know yet is how much educators will want to pick these things up themselves versus simply asking someone to build them for them. There's a real accessibility problem here. A lot of this work currently lives in places like GitHub, which is an excellent tool for software developers and a fairly hostile one for everyone else. Making these primitives genuinely accessible to non-technical people is itself unsolved work. ## The web app is no longer the unit of value The lesson-plan-generator category is enormously crowded. That's partly because wrapping a language model in a form is relatively easy. Give someone a text box, a few dropdowns and a prompt, and you have a product that can generate a lesson plan, a unit, a handout, blah, blah, blah. In the school where I work, in my wife's teaching, with my co-founder, and all over LinkedIn, I'm watching educators build their own solutions. They're doing it directly in Claude and ChatGPT, rather than waiting for an ed-tech company to build the perfect workflow for them. The models are dynamic enough now that a motivated teacher can create something surprisingly tailored to their exact classroom in an afternoon. That's a meaningful shift. It suggests that the valuable thing isn't always the application. Sometimes it's the context and the building blocks around it. Give someone a sufficiently capable model, the right knowledge, and a few useful primitives, and they can assemble something that would have previously required an entire product team. That changes what we should be building. Instead of trying to own the interface where the work happens, perhaps we should make the underlying context available wherever the work happens. Instead of building another closed system, we can build shared public infrastructure. Curriculum knowledge that machines can actually use. School context that can move between tools. Capabilities that agents can discover and invoke. Data and knowledge represented in ways that are useful to both humans and machines. The goal isn't to make educators come to our application. It's to make the things they already need available to them — wherever they happen to work. Maybe you don't need another application. Maybe you need the right primitives, the right context, and the freedom to build the thing you actually need.