You’ve written an IEP. You’ve assessed a student’s processing profile: working memory capacity, attention span, where they lose the thread. You’ve taken a compound instruction (“solve for the missing number, show your work, and explain your reasoning”) and broken it into three separate steps because you know the student will process the first one and lose the rest.

You’ve built scaffolding: a graphic organizer for paragraph planning, a sentence starter, a visual checklist. You’ve watched for the moment the student doesn’t need it anymore and taken it away, because a scaffold that stays permanent builds dependency instead of capability.

You’ve written individualized success criteria. One student’s goal is a complete sentence. The student next to them is writing a paragraph with a topic sentence and supporting detail. Same assignment. Completely different definitions of done.

You already know how to design for AI. Nobody in the AI field has figured that out yet, but I think they will.

The mapping

Everything you do in a classroom maps directly to AI system design:

Task decomposition. A language model given four objectives in one prompt processes the first with full attention. Each one after that degrades. You already know how to handle this: break the compound instruction apart, let each piece finish before the next one starts.

Scaffolding. A coordinator pattern dispatches tasks to the model one at a time, collects results, and synthesizes them. The model never sees the full complexity. When the task is simple enough that the model handles it reliably, the scaffold can come off. Same principle as removing the graphic organizer.

Individualized criteria. Evaluating writing for “voice quality” and “structural integrity” in the same prompt produces a blended average. Separate them. Give each dimension its own criteria and its own definition of success. Run them independently. This is how IEP goals work: each goal has its own measurable objective, its own baseline, its own progress monitoring.

Progress monitoring. Each output gets assessed against specific criteria before the next task begins. Iterative evaluation loops. Continuous assessment against individualized goals. You already do this every day.

Why the field needs you

The question that drives good IEP design (“what does this student actually need to succeed at this task?”) is the same question that produces better AI architecture: what does this system require to succeed at this task? And the people who’ve been asking that question about human learners for years are exactly who the field needs.

I taught a self-contained 4/5 bridge class in Sunset Park, Brooklyn. Twelve students, every subject, every accommodation. The framework I built for AI governance came directly from that room. I call it accommodation design, and the reason it transfers is that the underlying problem is the same: a system with specific processing constraints receiving tasks designed for a different processing profile.

You’ve been doing this work for years. The system you’re designing for is a language model now instead of a student, but the diagnostic muscle is the same one you’ve been building since your first IEP meeting.