Building an AI teaching assistant that became part of a home educator’s week
Paideia is a lesson app for home educators that I researched, designed and built alone. In July its pilot tutor said it wasn’t worth it, and by October she was using it on most days; this is how it got from one to the other.
- Role
- Solo: research, design and build, with Expo, React Native, Supabase and the Claude API
- Pilot
- One household; the pilot tutor is my wife
- When
- February to November 2026
23 of 31
days she used it in October, after saying in July it wasn’t worth it
Home educators teach alone, and AI tools either plan their week or replace them
There is no department to compare notes with and no scheme of work to measure against, so a lesson that went badly has nowhere to go. Planners point forward at the week you intend to teach and tutoring apps teach the child directly, which leaves the person doing the teaching to work it out alone.
-
Planners
Point forward
At the week you intend to teach.
-
Direct-to-child tutors
Replace the adult
The child is taught directly.
-
Paideia
Coaches the adult
Reads what actually happened in a lesson, and coaches the person who taught it.
The pilot tutor had logged two years of lessons in a paper planner, and nothing ever read them back to her. Planning ahead would have meant telling a home educator what a week ought to contain, which is often part of why a family leaves school in the first place, so Paideia works from what actually happened instead.
- May Reading a planner photo, tested before any screens
- 28 July She says it isn’t worth it
- August A lesson logged in a sentence
- 2 September She says it is
- 21 September Agents attack its limits
- 30 September Keeping an idea takes a tap
- October Every idea she tried feeds the next
- 16 November A second household starts
Part one, May to September: finding what was worth building
Reading photos of the pilot tutor’s planner worked, but it recorded what was taught, not how it went
The whole idea rested on reading a photographed planner page, so I tested that in May before designing any screens, and in August she submitted 34 pages in 11 minutes, unaided. But a planner holds what was taught, and almost none of it said how the child did.
Lessons logged before the conversation arrived
She said it wasn’t worth it on 28 July, and I kept that verdict on record
I’d named the cost of photographing her planner as the biggest risk before the session, and her answer confirmed it. From then on the app asked her the same question, and the pass mark was yes.
“No… it’s not something you can just upload once. You’ve got to keep doing it… that was quite a lot.”
Checking every lesson was the real cost, so I cut it and let her simply say what happened
Cross-referencing each lesson against her planner was, in her words, “actually one of the big things that was adding a lot of time.” The same day, lessons the app was sure of could be accepted in one batch, and by the end of August she could tell the app about a lesson in a sentence.
By September her answer to “was it worth it?” had turned from no to yes
She answered yes in the app twice in September, and when she was shown a month of each child’s work grouped by theme, she said “this is the whole thing.”
No 28 July, in a moderated session
Yes 2 and 14 September, in the app
Part two, September to November: learning fast once she was using it
The coaching sat five taps from where she spent her time, so I put it in the reply
In September she called the coaching “hidden”. In the chat, the first words of a reply arrived in under a second in October tests, on the screen where she is already typing.
Before: the value came later, somewhere else
- Photograph the planner and wait while it is read.
- Check each lesson against the page.
- Find the coaching, five taps from the home screen.
After: the value is in the reply
- Tell it about the lesson in a sentence.
- Read ideas to try in the reply, on the same screen.
Ideas she found useful were lost after the lesson, so I made keeping one a tap on the reply
In July she’d called quick ideas for ten minutes before a lesson “really useful”, but nothing kept them. She has kept 195 from replies since, and when she said “I do look at that ideas list a lot”, starring shipped the same day.
I gave the chat every idea she’d tried and how it went, so the next ideas start from what worked
Since October the chat is given every idea she tried and how it went, and an idea that didn’t land isn’t suggested again unchanged. So the loop is short:
- Ask for ideas
- Try one in the lesson
- Say how it went
- Get better ones next time
I decided where the AI stops, and had agents attack those limits about 300 times
It assists rather than teaches, the child is never a user, and it never goes near a diagnosis, because the distance between noticing a pattern and implying one is shorter than it looks. Seven agents tried to break those rules about 300 times, and the attacks that got through became fixed refusals and a set of tests I re-run.
Results
It became part of her week in October, and stayed there into November
That’s a tool she said wasn’t worth it in July. She also says it has made her more considered in how she records her lessons, most recently in November.
One household gives me findings I trust and a sample I don’t, so I’m opening it up
I’ve been interviewing other home-educating families, and the first new household was set up on 16 November. Only other households can show which of these findings are about home education and which are about one family.
Next time I’d put the value in the moment of use from the first version
In June she asked to log lessons in a freeform, diary-like way, and I parked it while I worked on reading the planner. The conversation turned out to be where the value arrives, and it is the same principle as getting a free trial to its first result: the closer the value sits to the moment of use, the tighter the loop.