How to use ChatGPT or Claude for your child’s IEP

Why generic AI advice falls apart

Ask ChatGPT for help writing a social story, and it will write one. It will be competent, warm, and pitched at an imaginary eight-year-old who makes eye contact, reads at grade level, and is not currently dysregulated because the fire drill happened at 10:15.

That is not a failure of the model. It is a failure of context. The AI has no idea that your child is minimally speaking, that they use a device, that “be flexible” is the single hardest instruction in their world, or that the word surprise reliably ends the afternoon. You know all of it. The AI knows none of it, and it will not ask.

So you get output that is fine for somebody else’s child, and you either rewrite it yourself or abandon it. Most parents abandon it, conclude AI is overhyped for special education, and stop.

The six things an AI actually needs to know

Across the families we have watched do this well, the same six categories come up. If your prompt covers these, output quality changes immediately.

1. How they communicate

Speaking, minimally speaking, AAC device, sign, gestalt language processing, echolalia. Also: how they communicate when they are upset, which is often not how they communicate the rest of the time.

2. How they process information

Processing speed, working memory, how many steps they can hold at once, whether written or spoken instructions land better, whether visuals help or overwhelm.

3. What dysregulation looks like, and what precedes it

Not just “has meltdowns.” The early signals, the reliable triggers, and — most usefully — what actually helps once it has started, which is frequently the opposite of what a stranger would try.

4. Interests, in specific detail

“Likes trains” is weak. “Knows every stop on the Blue Line in order and will correct you” gives an AI something real to build with. Special interests are the single highest-leverage thing you can hand a model.

5. Goals that are actually being worked on

The two or three current IEP or ISP goals, in plain language. Not the whole document — the things this month is about.

6. Safety and non-negotiables

Elopement risk, food restrictions, allergies, medication timing that affects the afternoon, topics that must not be raised. This is the category people leave out and then get burned by.

A prompt you can copy

Fill in the brackets and paste this at the start of a new chat. It is deliberately blunt — models follow direct instructions about tone better than they follow polite hints.

The last rule matters more than it looks. Without it, models fill gaps with assumptions about a typical child, and you will not always notice which parts were invented.

What never to paste into a chatbot

Consumer AI tools are not covered entities, and depending on your settings your conversations may be retained or reviewed. Assume anything you type could be read by a person at some point, and leave out:

• Full legal names — a first name or an initial is plenty for the AI to write well.
• Your address, school name, or district unless you have a specific reason.
• Medical record numbers, insurance IDs, dates of birth.
• Whole scanned IEPs uploaded without reading them first. They usually carry names, addresses, staff names and signatures on pages you were not thinking about.
• Other children’s names — siblings, classmates, the child in the incident report.

None of this stops the AI from helping. A model does not need your child’s legal name to write a good visual schedule; it needs to know they can hold two steps and love the Blue Line.

What AI is genuinely good and bad at here

Genuinely good

• Rewriting the same instruction eleven different ways until one lands.
• Turning a goal into a social story, a visual schedule outline, or a script for a hard transition.
• Translating clinical language into something you can say to a grandparent or a substitute teacher.
• Drafting the email to the school team when you are too angry to write it yourself.
• Preparing you for a meeting: what to ask, what a phrase in the document actually means.

Genuinely bad

• Anything diagnostic. It will produce confident text with nothing behind it.
• Legal advice about your rights under IDEA. Rules vary by state and change; a model will not reliably know which version it is describing.
• Judging whether a goal is appropriate. It has never met your child.
• Crisis situations. It is not a person and it is not on call.

The part nobody warns you about: doing it again

You will write that context block once and it will work. Then the chat gets long and the model drifts. So you start a new one — and the context is gone. Then you try a different tool, and it is gone again. Then the school gives your child a Chromebook with a different assistant on it, and you are typing the same six categories for the fourth time this month, from memory, at 11pm, leaving things out because you are tired.

This is the actual problem. Not that AI cannot help a child with an IEP — it can, considerably — but that the context which makes it help has to be rebuilt from scratch every single time, by the one person who already has the least time.

That is the problem SpanNova was built for. You answer the questions once, or hand it the IEP you already have, and it produces a structured profile file covering all six categories. You keep the file. You paste it into whatever tool you are using, and the explaining is done.

You do not need SpanNova to use anything on this page — copy the prompt above and go. But if you find yourself writing it for the fourth time, that is what we are for.