LLMs need limits: teaching students that AI does more with less

For all of last school year and much of this one, I gave students domesticated AI.

I created custom bots and then presented them to students so the bots could guide them through various experiences: professional moral dilemmas, potential AI impacts on their career, long-term project design.

Then, in the late fall of this year, I realized students were getting lots of support from those experiences, but they weren’t learning how to effectively interact with frontier AI models.

They hadn’t encountered AI in the wild.

So I created a bot with all the usual school-specific safeguards under the hood, but which otherwise acted like a frontier AI. It didn’t give any instructions or guide students through a predetermined pathway. In other words, it gave students the experience of looking at a blank prompt window and deciding what to type.

I then taught some basic prompting techniques, including giving the AI a role, saying exactly what you need, and then using iterative prompting to refine the output.

More importantly, though, I taught some basics of context engineering: explaining exactly what situation the student was in, uploading related assignments or other curricular materials, and when relevant providing cut and pasted examples of their work. Some students uploaded photos of handwritten materials.

In teaching both prompt engineering and context engineering, it became clear to me that students — through no fault of their own — treat AI like Google, in the sense that they just write in the exact question that’s on their mind, without any other context or detail. There was not much subtlety in students’ approach.

Digital natives, like all native speakers, don’t automatically understand the grammar underlying the tools they use.

So I was happy to see students refining their prompts over time, isolating the AI’s misconceptions or simplifications, and providing more and more robust and strategic context.

But coming back from winter break, it’s dawning on me that I haven’t yet taught a key tension in AI use, partly because it’s one I’m still learning myself.

Wide-open AI

I’ve been as eager as anyone to have students explore the farthest reaches of AI’s “thinking,” in part because that allows them to get quick access to career-related information that I can’t provide nearly as fast on my own.

Malik, a 12th grader in my career-connected learning program, is hoping to become an architect and is currently sketching some designs for a house. I ideally want him to speak with an architect to get feedback on his work. Until that happens, though, frontier AI can give him much better insights than I can.

Today at the start of class Malik uploaded a photo of his sketch to our AI, which told him he needed to add a vanishing point to his sketch to make it more effective. When I came back around to his desk later the period, he was midway through his revised sketch, with a vanishing point prominently featured in the middle. I asked him why that was necessary, and he explained it helped maintain realistic proportions.

That’s a case when wide-open AI was useful, because it could access specific domains of knowledge that a specific student needed.

Often, though, I find that kind of wide-open AI experience to be as detrimental for students as it is empowering. For example, when students want help launching a new project, the AI can sometimes know too much.

Zenobia, an 11th grader, today wanted to start a school drive to gather supplies to support unhoused people in our city community. We went to AI first, which resulted in information overload. The AI listed several different types of drives she could undertake, along with multiple different local shelters and which kinds of drive would benefit each. The AI also managed to bring in both national best practices for this kind of drive as well as very specific questions around state laws regarding solicitation.

It was a lot, for both Zenobia and me. We ended up going out into the hallway to call one local shelter and ask them for their advice, which left us much clearer on how to begin.

I see Zenobia’s experience happening more frequently than Malik’s. When students bring AI an open-ended problem, it tends to give them everything they could possibly do — every option, every consideration, every possible next step — when what they needed was one place to start.

The solution, though, isn’t going back to bots. Bots have their uses to be sure, but students aren’t learning to use AI if they only use someone else’s AI tool. A Waymo can get you where you’re going, but you don’t know how to drive any better when you get there.

The solution I’m coming to in my practice is helping students learn to limit LLMs.

Liberating constraints

One of my high school English teachers once explained why it was easier to write an essay when given a prompt, rather than just starting from scratch. He called the essay prompt a “liberating constraint.”

Students need to know how to give AI liberating constraints.

Here’s how I’m planning to teach that skill, as five steps for students:

1. Don’t type anything yet!

2. Upload everything relevant: class materials, assignment instructions, your own work so far (PDFs, photos, or copy-and-pasted work). The more context you give AI, the better it can see your specific situation rather than giving generic advice.

3. Decide what ONE thing you need right now. Not “help me with my project” but something limited: “What’s a realistic first step I could take today?”

4. Now start typing, and build constraint into your question. Start with this: “In your thinking and answers, give priority to what I’ve uploaded. Only consider external information once you’ve considered what I gave you.” Then ask your question or say what you need.

5. If AI still gives you too much, push back. Say something like: “That’s too many options. Review what I uploaded again and give me 2-3 ideas.”

The key step here, practically and conceptually, is #4. Students need to understand that the purpose of all the uploading, all the context engineering, is to create limits.

The AI needs to stay inside the fence students create, where it can do its best work. It needs to be able to look through that fence and bring in information from the wider world, but only as benefits its work inside that fence. This combines the wild and domesticated experiences into one that ideally provides the benefit of both.

Students will hopefully see that, to paraphrase Hamlet, AI does better bounded in a nutshell than when it makes them kings of infinite space.

I also hope there are two implicit lessons here:

Students can take charge of AI. They don’t need to settle for being passengers in this car.

Students’ intellectual and imaginative horizons are the point, because they are vaster than any space tokens can fill. That’s where we want them running wild.