I Used AI to Help My Daughter With Reading. Here's What I Had to Build Around It.

My daughter is 10, fifth grade, and about eight months out from middle school.

In September her reading assessment came back at 425L. For context, a fifth grader heading into fall should be somewhere around 700-800L. That gap was not small.

I'm not a reading specialist. I'm a dad who pays attention. So I did what a lot of parents do: I started researching, talked to her teacher, and eventually started wondering whether AI could help me build something more structured than flashcards and hoping. It could. But not in the way I expected.

The first thing I learned: AI without guardrails is just confident noise

The first time I asked Claude to help my daughter improve her reading, it gave me a perfectly reasonable, completely generic response. Fluency activities. Vocabulary building. Read together. It sounded good. Could have applied to any kid, any problem, any grade.

That's the trap with AI in education. The output sounds authoritative. It uses the right vocabulary. And if you don't know what you're looking for, you'll take it and run with it.

I needed something more specific. And I needed a way to make sure that specificity was grounded in something real, not just plausible-sounding.

So I did two things before I built anything.

First, I sat down with her teacher and asked what she was actually seeing in class. She told me my daughter had specific gaps: cause-and-effect reasoning, tracking plot across a full text, and using evidence from what she'd read to support an answer. She also mentioned that she tends to wait in class. She'll hold back until a peer starts before she commits to an answer. That detail mattered later.

Second, I learned enough about reading science to ask better questions. I'm not an expert, but I now know Scarborough's Reading Rope. I know what decoding versus language comprehension means. Her decoding is actually fine -- she reads the words. The gap is in the thinking skills that sit underneath comprehension.

With that as my foundation, I went back to Claude and built it a role.

Building the role

I wrote a system prompt -- a set of instructions that lives at the top of every conversation I have with Claude about her reading. It tells Claude who it is (a reading specialist and instructional designer, not a tutor), what her specific gaps are, what the session structure looks like, and what the rules are.

The rules are the whole thing.

I locked in: sessions cap at 20 minutes. She attempts every question before she gets help. There are only specific allowed ways to offer a hint. Difficulty doesn't go up until independence goes up first.

None of these came from me. The hint structure is based on gradual release -- a real instructional framework. The difficulty rule is basic mastery-based progression: you don't advance the challenge until the skill is stable where it is. I found these, cross-checked them against what the teacher was telling me, and baked them into the instructions.

What I ended up with is Claude functioning as an instructional designer. It generates story packets for our sessions -- four pages, custom characters she actually cares about, cause-and-effect relationships embedded in the plot. It gives me think-aloud scripts so I know what to model and when to stop talking. And it helps me interpret what I observe across sessions, because I report back after each one and the system responds to that data.

It doesn't teach her. I do. What it gives me is structure and a thinking partner who has read everything I've fed it about how this specific kid learns.

Verifying with the teacher

Here's the part I'd want any other parent to hear: I take notes to the teacher.

After clusters of sessions, I write up what I observed. What she got on her own. Where she got stuck. What patterns are showing up. And then I ask the teacher whether I'm reading it right.

Twice, she caught something I had wrong. Early on, I thought my daughter's habit of answering from memory instead of going back to the text was a confidence issue. The teacher told me no -- it's the exact same pattern she sees in class. She leans on what she already knows instead of returning to the source. That's a behavior pattern, not an anxiety response. It changed how I designed the next several sessions.

I also asked whether my Lexile targets were calibrated right. The teacher told me she doesn't use Lexile for instructional decisions. She uses Fountas and Pinnell levels. If I hadn't asked, I'd have been optimizing for the wrong metric.

The teacher can't run an intervention with every kid after school. But she can be the quality check on what a parent is doing at home. That feedback loop is what keeps AI-generated content connected to reality.

The results, honestly

Her reading score in September was 425L. By January -- before we started the structured sessions -- it had climbed to 675L. That growth happened through the school year, not through this intervention. The January number is our baseline, not a win.

Since starting in January, we've run through 21 custom story sessions. Four sessions a week, 20 minutes each. The skills we're targeting are moving, but slowly.

Cause-and-effect reasoning is getting more consistent. She now produces both the cause and the effect reliably when I prompt her, after months where she would only give me one side. Text evidence is still emerging. She'll find evidence when I ask her to show me where in the story, but she doesn't yet go back to the text automatically before answering. That's the current edge we're working.

What I can say truthfully is that the shape of her effort has changed. She used to wait for me to rescue her when a question got hard. Now she pauses and tries. She tells me directly when a session feels too heavy -- that kind of self-advocacy didn't exist in October. And she caught a design error in one of the packets a few weeks ago. A character was named in a question but never appeared in the story. She flagged it. That's active comprehension monitoring. A few months ago, she wouldn't have noticed or wouldn't have said anything.

Her March STAR score came back recently. I'm looking at it alongside the teacher before I interpret it myself. That's the rule I set for myself early on and I'm sticking to it.

Progress is real. It's slow. Middle school isn't waiting.

What you can try (guardrails included)

Start with the teacher, not the AI. Build a role with rules, not a prompt -- specific gaps, time limit, fallback plan, attempt-first. Five to ten seconds of silence before you step in is longer than it feels. Don't fill it. Report back after clusters of sessions and ask whether what you're seeing matches what she sees. Go slower than feels right. Independence is when they do it before you ask, not when they do it with you watching.

The AI isn't doing the teaching. You are.