AI is Being Built for Coders
The reckoning for those of us in the humanities was inevitable
A slew of frontier model updates and new capabilities have recently dropped in quick succession.1 One, Claude Design, got lost in the shuffle. While playing around with it over the past week, something became clear. Now that code generation rather than text is the leading edge of AI capability, it comes with a cost. AI is getting more complicated to use. Vibe coding was a thing a year ago2 - non-coders using natural language to make AI write code for them. What’s emerging now is something new: native interfaces specifically tailored to create things - a document, a presentation, an app, a web tool. Most people are not taking advantage. The barrier to entry to really leverage what the most advanced AIs are capable of now requires thinking like a coder.3
Unfortunately, I’ve never been a coder. Most teachers reading this probably aren’t either. An entirely new vocabulary, one I've been learning mostly through newsletters geared toward entrepreneurs, is leaking into everyday conversation: APIs, tokens, context windows, agents, MCPs, harnesses, CLIs. You can’t read deeply in the AI space now without encountering hyperventilating claims about new workflows that produce all sorts of new digital artifacts and prototypes. Software engineers, marketers, and designers are rapidly absorbing these new tools into their jobs. That’s the hype. But hype doesn’t mean things aren’t actually changing. My experience is ahead of most teachers', but the wall is real - and humanities teachers will hit it hardest.
For most of my professional life I knew, in some abstract way, that everything I typed and clicked online was built by someone in a backroom who knew how to make it work. That was the deal humanities educators made with the digital age. We taught our students how to read critically, write cogently, and think deeply, all while seamlessly using digital tools - the internet, computers, and software. The people who knew the secret language of the machines handled the machines. All that happened somewhere else.
That deal is changing.
The original promise of generative AI was democratization through natural language. Anyone able to type can access expertise. That was true for the first wave of chatbots and it’s still true today. A teacher can quickly draft a rubric or a lesson plan. Someone who knows nothing about auto mechanics can teach themselves rudimentary maintenance tasks. AI quickly became an interactive YouTube tutorial. The friction between having a question and getting a coherent answer collapsed almost overnight.
But the simple chatbot is now the floor. The most powerful uses of AI sit several levels above where most people are currently working, which means we are drifting back toward a real expertise gap. To build something really useful with Claude Design - or Cowork, or Codex, or any of the platforms sprouting up everywhere - you have to think like someone who knows what a digital artifact really is, when to deploy versus refine, how to prompt at a granular level of specificity, and when to ask for code rather than revision. Hardest of all: how to recognize when the AI is offering you a design choice instead of a final product. A mistaken entry has consequences - in token use, in time spent waiting for an output, and in the rate limits you'll hit faster than you expect.
You don’t have to write any code. You do have to think the way coders think, make the kinds of decisions they make, and understand the language they use. A new threshold needs to be crossed.
Models, Apps, and Harnesses
If you want to see the issue up close, read Ethan Mollick’s latest post on GPT-5.5. Mollick is one of the most accessible writers about AI and he works hard to translate for non-technical readers. But even his current framing includes assumptions about what general readers are tracking. If you haven’t kept up, his posts are liable to be overwhelming.
He encourages us to think about AI as three interlinked concepts: models, apps, and harnesses. Models are things like Opus 4.7 and GPT-5.5. Apps are the products that let models do real work - claude.ai, chatgpt.com, Claude Code, and Codex. Harnesses are the tools the AI uses and how it is hooked up to them. That last one is probably the deal breaker for most teachers. How many educators do you think are linking APIs to GitHub? Setting up MCP servers? Are you still with me? This is where the real power resides for serious AI users, and it is currently beyond the reach and expertise of most educators.
Things are moving so fast that even the most vigilant cannot keep up. That is becoming more true by the week. A humanities teacher reading the GPT-5.5 piece is not reading about whether students can cheat on an essay. Mollick describes generating a near-PhD-quality academic paper from four prompts using Codex and an interactive 3D simulation of an evolving town across 6,000 years from a single prompt. These are now ordinary outputs from someone who knows how to integrate models, apps, and harnesses into a repeatable workflow.
Even Mollick has acknowledged this gap. In a post from late March, writing about the most powerful tools currently available, he admitted:
These tools are terrific, but they are really built for programmers. They assume you know Python and Git. Their interfaces look like a 1980s computer lab. For the 99% of knowledge workers who are not developers, these powerful AI tools are not optimized for them.
Tools like Claude Design are the industry's attempt to fix this. The capability is there but what is new is the scaffolding meant to let non-technical users reach it. The catch, based on my experience, is that the new scaffolding is itself technical. The interface is better, but the conceptual demands are still out of reach for most. Maybe that will change, but right now it’s an impediment.
Hitting A Wall
Last December I appeared on the leaderboard of a brand new platform called AI Cred4, and the founder’s follow-up post name checked me as one of the surprises - a 30-year history teacher with no coding background showing up alongside engineers and consultants. In February I wrote that Claude Cowork “just worked” for me, and I meant it. Building online tools took some figuring out, but I did it. I suspect I'll get there eventually with Claude Design, but I'm finding myself out of my depth more often than I once was.
I need to reread every sentence twice. I’m having flashbacks to my days fumbling through FORTRAN and Pascal in the 80s. The skills that got me on that leaderboard last fall haven’t gone away. The problem is everything that’s piled on top of them since. Keeping up with the models and new releases is practically a full-time job.
The tantalizing thing is that I do see the value. I understand why these tools are a breakthrough and eventually worth learning. But taking the time to master them is an opportunity cost I can’t currently spend. It all makes me nervous about what comes next.
Last week I asked Claude Design to revise a one-page professional summary I’ve used for years. The draft came back elegant and the formatting impeccable. I was inserting updates from the last few years and needed to know how best to export the file. It recommended a standalone HTML file - not PDF and not Word. Then it offered an unsolicited note: “Start a new chat to save 107k tokens of context. This uses your rate limits more effectively.”
I knew what “standalone HTML” meant because of the Cowork projects I've done. I also knew that saving “107k tokens of context” was a good thing. A teacher who recently upgraded to Claude Pro to try Claude Design might not have known what either phrase meant.
It is not that learning these capabilities is hard, exactly. You can ask the AI when you get stuck and it will generally walk you through it. What the companies are doing, however, is putting the newer models out of reach for all but the most dedicated users with the resources and time necessary to devote to it. It’s understandable why companies will invest in this. It’s not yet clear why schools should.
The Eternal Skills Question
The best argument against using AI in classrooms goes like this. Teach students the eternal skills - writing, judgment, critical thinking, the ability to refine ideas - and AI fluency will follow. Kids need to know how to think before ever touching a chatbot. Concrete elements of AI instruction, like prompting, can be learned later. I tend to agree with all of this in the abstract. The strongest version of the argument is that schools exist to build durable human capacities, not to chase whatever the workforce happens to need this year.
But the eternal skills argument assumes students will not be using AI during that time. That is impossible unless we lock them in a room without technology. Some parents would willingly make that tradeoff. Most schools cannot.
It also assumes the AI students encounter at twenty-three will resemble the AI they are using at fifteen. It won’t. Students graduating this May and June will matriculate to college and eventually workplaces where the entry-level fluency expectation is already past the chatbot tier. The AI of 2034 will be unrecognizable to them, and the judgment required to use it well is not something that can be snapped on at the end of college. AI is the first tool that talks back in any meaningful sense - it argues, flatters, redirects, and sometimes deceives. Nothing in our experience tells us how long it takes to develop fluency with something like that. Withholding AI from students until they’re “ready” assumes a static technology which currently doesn’t exist.
I'm not convinced the eternal skills will be enough on their own, or at least, they are going to have to be developed alongside the reality of student use. AI literacy is on track to become a baseline professional expectation and the cost of being late is likely to compound quickly.
Class of 2032
My middle-school daughter graduates high school in 2032. Even if you accept the argument that high school is not workplace training, not exposing her to AI and what it is capable of during her high school years feels like an abdication of our educational responsibilities. We know they're using it already - every conversation I've had with actual students confirms it. Teachers cannot use AI themselves while refusing to explain it to them. The question no one has been able to answer yet is how to do that well. Can we learn this with them rather than ahead of them? The challenge is how to teach a tool we’re still figuring out ourselves.
This is uncomfortable territory because it implies that the humanities teachers leaning into the AI conversation in schools - and I am one of them - need to learn things we did not sign up to learn. Not exactly how to code, but how to adopt the mindset and understand what computer code is. We need to recognize when a tool should be used as an artifact builder rather than a chatbot and, most importantly, when it should not be used at all. Even more daunting is we may need to adapt the vocabulary of the software engineer to the vocabulary of the classroom.
The Home Stretch
As we head into the home stretch of the school year, I have been reflecting on what has changed for me over the past twelve months. “Everyone is Cheating Their Way Through College” was published on May 7th, 2025. That is old news now, and a tired take, though I do not know exactly how much has actually changed in classrooms. There may yet be a wave of success stories about blue books and oral exams. My experience has been mixed.
The real story for me is how I have moved from a purely text-based use of AI to one predicated on the ability to generate code to do things. Yes, it is partly the “agentic layer” that was promised last year and is now taking root. But the more important point is that it is going to require an entirely new set of skills for non-coders to do the kinds of things that come naturally to coders.
When I spent a summer as a paralegal, one of the partners explained to me that every profession has its own secret vocabulary, its own language that sets it apart. Generative AI gave us, briefly, the illusion that we could keep swimming solely in the world of words we understood. The truth is we have been living in someone else's domain all along. We just did not have to see it. Humanities teachers may have to learn to read in a language we did not train for, if we intend to keep up.
Connect With Me
Beyond this newsletter, I work directly with schools, educators, and other institutions about pedagogical questions raised by AI. Take a look at my website and reach out - I’d love to hear what you’re working on.
Opus 4.7, Claude Design, OpenAI’s Image 2, and now GPT-5.5 all landed inside an eight-day stretch in April 2026: Anthropic released Claude Opus 4.7 and the Claude Design tool on April 16; OpenAI released ChatGPT Images 2.0 (model name gpt-image-2) on April 21; and OpenAI released GPT-5.5 on April 23. This list doesn’t include Anthropic’s Claude Mythos Preview - its most capable frontier model, held back from general release and instead distributed to a small group of cybersecurity partners through Project Glasswing, which was announced on April 7. Anthropic on Opus 4.7; OpenAI on ChatGPT Images 2.0; OpenAI on GPT-5.5; Anthropic on Project Glasswing / Mythos.
Andrej Karpathy, who coined “vibe coding” in February 2025, retired the term in February 2026 in favor of “agentic engineering” because the workflow now demands “more oversight and scrutiny” than the original vibe-coding ethos allowed. Vibe coding is passé. Karpathy has a new name for the future of software. The New Stack, February 10, 2026.
Perhaps more significantly, it requires money. Claude Code, Cowork, and Design are only available on paid plans. Codex is currently included with the free ChatGPT plan, but on a “limited time” trial basis and with heavy usage caps.
AI Cred (aicred.ai) scores users on actual AI fluency rather than self-reported familiarity.
Hi — I taught math in NYC schools and now private tutor while building an AI math tutor for my students. Your post touches on real struggles teachers are about to face, and I agree with most of it. I'd push back on one piece though.
For me this hasn't really been about learning to "think like a coder." It's been about tolerating being wrong over and over while you figure out what actually works with AI. You build, you test, you realize you were thinking about it wrong, you start over. That's a teacher skill — we do it with lessons all the time.
What worries me more than the coder gap is the time gap. Teachers don't have evenings and weekends to spend on this the way solo builders do. That's the real wall, and I don't think anyone in the industry has answered it yet.
All well-documented and, as usual, insightful!
Yet I'm curious about the lens of "humanities teachers needing to get on board," as it feels like a much broader expertise gap—one that, at least for me, feels more and more divorced from the parameters of a typical high school English classroom.
Yes: I think we should continue to be offering opportunities for digital literacy, introductions into coding, etc. If I had my way, this would be a core sequence through high school alongside Math, Science, ELA, and Social Studies.
But: I guess I don't feel obligated at all to set aside time as an ELA teacher to prioritize this right now, particularly when it comes to the coding language/requirements for these new tiers of use.