We don’t believe kids should learn to prompt.

Why prompting looks like a teachable skill.

You can see the appeal. Prompt engineering has all the surface features of a curriculum-friendly skill. It’s specific: you write a sentence, the model returns a thing, you can grade the thing. It’s measurable: better prompts produce better outputs, by some criterion. It’s teachable: you can hand a kid a list of patterns (be specific, give examples, ask for step-by-step) and they will produce better outputs within a single sitting. A school administrator can write “Prompt Engineering 101” on a syllabus and parents nod.

It also feels productive in a way that’s rare for AI education. Most other ways into the topic ask the kid to confront something hard: how a transformer works, what a training corpus is, why a model says confident wrong things. Prompting feels like a real skill you can practice. The instant feedback loop is genuinely satisfying. We’ve watched kids work on a single prompt for twenty minutes, iterating, getting closer, finally getting the thing they wanted. They feel like they learned something. In a narrow sense, they did.

So the question isn’t whether prompt practice has value. It does, in the moment. The question is whether it’s the right thing to organize a kid’s AI education around. Our answer is no, for two reasons that compound: the skill doesn’t age well, and it isn’t what the research literature calls AI literacy in the first place.

The skill is transient.

Look at what changed between 2022 and 2026. In 2022, a working prompt-engineer’s toolkit had specific incantations. Let’s think step by step. You are an expert in X. I will tip you $200 if you do well. These patterns produced real, measurable lifts on the models of that era. Today, the frontier models route around almost all of them. The step-by-step trick was absorbed into the default behavior. The expert-roleplay trick became unnecessary. The tipping trick was always a hack, and it’s now a curiosity.

The pattern is reliable: a prompting technique that gives you a real edge becomes table stakes within a model generation, then becomes obsolete by the generation after. The half-life of any specific prompt pattern is now roughly twelve to eighteen months. A kid who carefully learns the 2026 prompt-engineering canon will find, by the time they’re in high school, that most of it has aged out. The model is doing the work the prompt used to do.

This isn’t hypothetical. It’s the same pattern that played out with search-engine optimization in the 2000s. Early SEO was a real craft with specific levers (keyword density, meta tags, link-farming) that produced measurable wins. The search engines absorbed each lever as a signal and then routed around the ones that became gameable. The people who built careers around the 2005 SEO playbook had to rebuild their entire skillset twice by 2015. The durable skill turned out to be something else: knowing what your reader actually needed.

Andy Matuschak has written about the difference between durable and ephemeral skills in learning more broadly: skills tied to a specific tool’s current behavior degrade as the tool changes; skills tied to the underlying activity persist.¹ Writing persists. Typing on a Selectric does not. Reading persists. Looking up something in a card catalog does not. The same shape applies here. Working with AI persists. The specific phrasing tricks that get the 2026 model to behave do not.

What AI literacy actually is.

The most-cited piece of work in this field is Duri Long and Brian Magerko’s 2020 CHI paper, What is AI Literacy? Competencies and Design Considerations.² The paper is the result of a systematic review of the AI-education literature, and it lands on a set of competencies that together define what it means for a non-specialist to be literate in AI. We’ve cited it elsewhere; this is the post where it matters most.

The framework names sixteen specific competencies, organized around five core questions. Here is the five-question version, paraphrased to fit a parent reading this post:

1. What is AI? Recognizing AI when you encounter it. Knowing what it can and can’t currently do. Distinguishing general AI from narrow AI.
2. What can AI do? Knowing the kinds of problems AI is suited to (and the kinds it isn’t). Understanding strengths and weaknesses across domains.
3. How does AI work? Some working theory of representations, decision-making, machine learning steps, and how human knowledge gets into the system.
4. How should AI be used? Critical evaluation. Understanding programmability. Awareness of bias and ethical issues.
5. How do people perceive AI? Knowing that AI is designed by humans, that those humans had goals, and that the kid is one of the people who can build it.

Read the list. Notice what isn’t there. Nowhere in Long and Magerko’s sixteen competencies does “prompt engineering” appear. Not as a sub-competency. Not as a recommended pedagogical entry point. The closest thing in the framework is competency 12 (“programmability”), which is about understanding that you can shape an AI’s behavior through inputs — a much broader claim than “learn this prompt pattern.”

This isn’t Long and Magerko alone. Stefania Druga’s decade of work on Cognimates at MIT lands in the same place: the durable AI-literacy goal is helping kids build correct mental models of what a model is doing, not training them to operate today’s interface.³ Mitchel Resnick’s tinkering-versus-planning distinction, developed across his work at the MIT Lifelong Kindergarten group, points the same direction: the durable skill is the iteration habit, not the up-front specification.⁴

If the people who’ve been studying this for fifteen years don’t name prompt engineering as the skill, why is the rest of the industry organizing around it? Because it’s easier to sell.

What we teach instead.

Inside the studio, three habits get built session after session. None of them is prompt phrasing. All three map onto Long and Magerko’s framework.

A working theory of model failure. Every time the studio’s AI partner proposes a change, the kid sees what it’s about to do before it does it. The tool-trace pane shows each step. The ChangeDisclosure card shows the diff in plain English. When the model gets something wrong (it will), the kid can see where the wrongness started. Over weeks, the kid develops intuitions: this kind of request the model handles well, this kind it confabulates, this kind it does mostly-right but with a specific failure mode. That’s a working theory. It’s the thing we’ve written about in when AI is wrong and it’s the foundation of competency 3 in the Long & Magerko framework.

The durable skill is deciding what to keep, not getting the model to say the right thing on the first try. On where AI literacy actually lives

The habit of deciding, not accepting. The studio architecture forces the kid to make a decision on every AI proposal: Keep, Review, or Undo. There’s no “the AI did it for me” mode. By the kid’s tenth wizard, they’re evaluating proposals fluently. By the fiftieth, they have taste. This maps onto Long & Magerko’s competency 7 (critical interpretation of AI output) and competency 11 (human-AI collaboration). It also maps onto Yasmin Kafai and Quinn Burke’s argument in Connected Code that the difference between consumers and makers of digital media is the practice of deciding what stays.⁵

An iteration loop. Real projects don’t ship at v1. The kid’s first wizard rarely produces the thing they wanted. The second one is closer. The fifth one is right. The studio celebrates v5s and v7s on the family dashboard because iteration is the move, and the move is what we’re trying to build. This is Resnick’s tinkering, restated for AI. It’s also competency 6 in the Long & Magerko framework (representations and decision-making), where the kid develops a feel for what the model is choosing among.

None of these three habits depends on a specific prompt pattern. None of them goes obsolete when the next model ships. All three transfer to whatever AI tool the kid is using in 2030, when the studio’s 2026 prompt patterns will be three model generations stale.

Typing is not the writing.

The honest framing is this. Prompt engineering is to AI literacy what typing is to writing. Useful skill, lower in the stack, worth picking up by osmosis. Not the thing you organize a curriculum around. You don’t hand an eight-year-old a typing tutor and call that English class. You hand them a story they want to tell and let the typing emerge from the wanting.

Same shape here. The kid in the studio types prompts. They get better at it because they’re doing it. We don’t teach the prompts, because the prompts will change. We build the loop around the durable thing: a kid who watches the model work, decides what stays, and iterates until the artifact is theirs. That’s the kid we want to graduate. The 2026 prompt patterns are a footnote. The competencies are the skill.

If you want the deeper framework, our breakdown of the Long & Magerko competencies is at AI literacy: what it actually means. If you want to see what deciding-not-accepting looks like in product form, the demo at tellandshow.ai/try takes about three minutes. And if you want the broader argument about why we build the studio the way we do, start with why making is learning.