A kid handed in a ChatGPT essay word for word. He wasn't "lazy."

A colleague and I were talking about learning design. I showed her Clief Notes.

She mentioned her son and daughter both have dysgraphia. Writing is physically hard, slow, and exhausting for both of them. Their ideas are good. Getting them on a page is the problem.

Then she said something that stuck. Her son had recently handed in a ChatGPT-written assessment without changing a single word. Not because he was lazy. Because the AI just did it for him, and he didn't know how to make it his own.

She asked if I could build something different.

Something that would help him do his own work without doing the work for him.

So I built it.

A Claude system prompt, not a folder structure yet, but a detailed one, with 10 task types: analytical writing, annotation, creative writing, short answer, practical reports, maths, recall, brainstorm, emails, and "not sure what this is." Each type gets a different level of scaffolding.

Her daughter needed a different version entirely. Different year level, different profile, different adjustments.

So now there are two.

The first thing built in: voice. Writing is the barrier for them, not thinking. So when their typed answers start getting short, Claude notices and prompts them to switch. Talk instead of type. The ideas come out the same way, just through a different door.

---

The rules that came from watching real sessions break down:

Claude cannot start from nothing. The student has to give it something first, even three messy sentences. Claude can only extend, restructure, or shape what is already there. The seed has to be theirs. And if Claude adds anything interpretive that the student didn't generate, it has to say so out loud: "I added this part, keep it, change it, or cut it?" The work has to be defensible as theirs. That is the whole point.

Don't ask "happy with that?" after every sentence. Just keep moving. One real check per paragraph at the end: "Read it through. Does it sound like you?" The first version asked this constantly. By the third paragraph she was typing "happy" without reading it.

If they paste the same thing twice, don't point it out. Take the second paste as their current answer and keep going. They're probably tired.

Don't celebrate every turn. "Good" or "yep" is enough. Save "perfect" for when it genuinely deserves it. Constant praise stops landing.

Don't teach new terminology mid-task. Finish the work first. Vocabulary comes after.

Label every element as required or optional. Don't make a student who takes things literally guess which parts count.

The four interrogation questions only run for graded analytical writing. Never for creative work or annotation. They kill it.

At the end: name what's missing once, then ask, add it, leave it, or come back tomorrow? The student decides when they're done. Not the AI.

---

We've been iterating based on real session data. Chat exports, actual tasks, rubric analysis. Both profiles are on their second and third versions. Still imperfect. Still improving.

The next step in my head: if a student had approval to record lessons, that audio could be transcribed and live in the project as context. Claude would actually know what the teacher said in class, not just generic subject knowledge. A folder on the desktop. Subject folders. Lesson transcripts. One specialist per task type. Built around how that specific student learns. For kids with dysgraphia, this matters more than most. Taking notes by hand while trying to listen is exactly the kind of thing that breaks down for them. Recording solves that.

I don't know if that's a product or a school policy nightmare. Probably both.

And it doesn't have to stop at dysgraphia. ADHD, processing differences, executive function challenges. Kids who have good ideas and struggle to get them out.

I'm not an expert in children's education or learning support. I'm a adult educator/digital marketer who had a conversation that led somewhere unexpected.

So I'm opening the floor. If you work in education, occupational therapy, learning design, or have a kid with learning differences yourself, what am I missing? What would make this actually useful at scale in a school context? Especially where there is a lack of funding.

Happy to share the prompts with identifying details removed if there's interest. Drop a comment.

8 comments