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

My kid is in elementary school, not far out from middle school. The first reading assessment of the year came back at 425L. The benchmark for that grade heading into fall sits somewhere around 700 to 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 the 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 kid improve their 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 the teacher and asked what was actually showing up in class. My kid had specific gaps: cause-and-effect reasoning, tracking plot across a full text, and using evidence from the reading to support an answer. There was also a tendency to wait. My kid holds back until a peer starts before committing 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. The decoding is actually fine. My kid 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 this. It tells Claude who it is (a reading specialist and instructional designer, not a tutor), what the 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. My kid attempts every question before getting 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 my kid 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 my kid. 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 my kid got alone. Where the sticking points were. What patterns are showing up. And then I ask the teacher whether I'm reading it right.

Twice, the teacher caught something I had wrong. Early on, I thought the 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 that shows up in class. My kid leans on what's already known 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 Lexile isn't what drives instructional decisions in that classroom. Fountas and Pinnell levels are. If I hadn't asked, I'd have been optimizing for the wrong metric.

A teacher can't run an intervention with every kid after school. But a teacher 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

The first score was 425L. By the next assessment, before we started the structured sessions, it had climbed to 675L. That growth happened through the school year, not through this intervention. The 675 is our baseline, not a win.

Since starting, 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. My kid now produces both the cause and the effect reliably when I prompt for them, after months of getting only one side. Text evidence is still emerging. My kid will find evidence when I ask where in the story, but 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 the effort has changed. My kid used to wait for me to rescue when a question got hard. Now there's a pause, then a try. My kid tells me directly when a session feels too heavy. That kind of self-advocacy didn't exist a few months ago. And my kid caught a design error in one of the packets a few weeks back. A character was named in a question but never appeared in the story. My kid flagged it. That's active comprehension monitoring. A few months ago it wouldn't have been noticed, or wouldn't have been said.

The most recent STAR score came back not long ago. 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 the teacher 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.My kid is in elementary school, not far out from middle school. The first reading assessment of the year came back at 425L. The benchmark for that grade heading into fall sits somewhere around 700 to 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 the 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 kid improve their 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 the teacher and asked what was actually showing up in class. My kid had specific gaps: cause-and-effect reasoning, tracking plot across a full text, and using evidence from the reading to support an answer. There was also a tendency to wait. My kid holds back until a peer starts before committing 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. The decoding is actually fine. My kid 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 this. It tells Claude who it is (a reading specialist and instructional designer, not a tutor), what the 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. My kid attempts every question before getting 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 my kid 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 my kid. 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 my kid got alone. Where the sticking points were. What patterns are showing up. And then I ask the teacher whether I'm reading it right.

Twice, the teacher caught something I had wrong. Early on, I thought the 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 that shows up in class. My kid leans on what's already known 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 Lexile isn't what drives instructional decisions in that classroom. Fountas and Pinnell levels are. If I hadn't asked, I'd have been optimizing for the wrong metric.

A teacher can't run an intervention with every kid after school. But a teacher 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

The first score was 425L. By the next assessment, before we started the structured sessions, it had climbed to 675L. That growth happened through the school year, not through this intervention. The 675 is our baseline, not a win.

Since starting, 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. My kid now produces both the cause and the effect reliably when I prompt for them, after months of getting only one side. Text evidence is still emerging. My kid will find evidence when I ask where in the story, but 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 the effort has changed. My kid used to wait for me to rescue when a question got hard. Now there's a pause, then a try. My kid tells me directly when a session feels too heavy. That kind of self-advocacy didn't exist a few months ago. And my kid caught a design error in one of the packets a few weeks back. A character was named in a question but never appeared in the story. My kid flagged it. That's active comprehension monitoring. A few months ago it wouldn't have been noticed, or wouldn't have been said.

The most recent STAR score came back not long ago. 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 the teacher 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.