Coaching the adult who teaches, rather than tutoring the child

Context

My wife home-educates our children, and for two years she had been logging each day's lessons in a paper planner. It held almost everything she knew about how the teaching was going, and nothing ever read it back to her.

Paideia started in February as a Claude project. I added transcripts of recorded lessons and asked whether any of it would help the person teaching. It did, which turned the question into how to make that a product, and that household became its only pilot.

The problem

Home educating is not a smaller version of school. There is no department to compare notes with, no scheme of work to measure against, and no external mark of whether a year went well. The tools that exist mostly assume otherwise.

• Working alone. No colleague to think out loud with, and in a home-educating context admitting that a lesson went badly carries a weight it would not carry in a staffroom.
• A record that only records. The planner captured each day faithfully and gave nothing back.
• No view of the arc. Nothing showed how a child had moved across a term or a year, which matters more now that the Children's Wellbeing and Schools Act requires every local authority to keep a register of children not in school.

Research and insight

The tools split cleanly into two bands. Planners point forward at the week you intend to teach, and direct-to-child tutors replace the adult altogether. Neither does anything for the person doing the teaching.

The February recordings added the finding the rest of the project kept returning to. Whether a child got an answer right says little on its own, because the useful part is the route there: how much prompting it took, which object on the table helped, and whether it held once the support was taken away.

Strategic reframe

She was already logging retrospectively, and using that logging as her reflection. Designing forward would have meant competing with planners and asserting what a week ought to contain, which in home education is the wrong posture, because the approach is tailored to the child by definition. So the tool reads what actually happened and coaches from that, which makes it empirical rather than prescriptive.

Extraction before screens

The whole product rested on one question, whether a photograph of a handwritten planner could become structured lessons, so I answered it before designing anything. In May the extraction read a real planner page with more than 90% accuracy, checked by hand, and the screens followed in June.

Her own corrections gave the wider picture. 24 of 28 pages needed none across 425 cells, and the 4 that failed, at 27 to 43%, were all the right-hand page of a spread. I had put it down to volume, and it was a missing header: those pages carry no child names.

The extraction worked, and it turned out to be reading the wrong thing. A planner records what was taught and cannot hold how the child did. As she put it in the second session, a line saying a child did three pages of work cannot tell the system whether they were three essays or three paragraphs.

The no, and what earned the yes

Each moderated session had accept criteria written down beforehand, so a disappointing result would be a finding rather than a matter of interpretation. The third, on 28 July, asked whether photographing her planner had been worth doing, and the answer was no.

It was not a rejection of the idea. In the same session she described what she would value, a picture of the weeks rather than a card after every lesson and quick ideas for when she was stuck, and she named what cost her most, which was approving every lesson one at a time.

So that is what I built next. On 2 September she was shown themes and patterns across a month for each child and subject, with a way to consolidate or expand any topic in it, and the answer arrived before the question was asked.

The same session said what was still wrong. The part she valued sat five taps from the home screen, and the synthesis spent its words handing her own facts back, which left her fact-checking it before she could trust the next line. Her fix became a rule: facts she gave come back as short lines with their source, and prose is spent only on what to try next.

Design decisions

Most of the decisions that mattered were about restraint rather than capability. The model can do considerably more than the product lets it do, and almost every rule below exists to stop it doing something it would otherwise do fluently and wrongly.

A teaching assistant, not a teacher

A tool that tells a home educator what she ought to be teaching is a red rag, because not being told what to do is often part of why a family is home educating in the first place. So Paideia sits as a teaching assistant rather than a teacher, a coach or a mentor. It can notice things and ask about them, and it can suggest a way to consolidate something she has taught, but it never decides what comes next.

The cost was the checking

The pilot showed the real cost was checking, because she had to approve every lesson against her planner, and she asked to see only the ones the system was unsure of. So lessons read with high confidence could be accepted in one opt-in batch, and the rest queued for her with the reason attached.

It helped less than it should have. On 3 August she submitted 34 pages in 11 minutes with no instruction and still went through them card by card. The conversation removed the queue instead, because her own words cannot be misread the way handwriting can, so nothing she tells it waits for approval.

A conversation of length one is complete

A surface which can ask can drift into one that expects an answer, so the floor is explicit: she can open it, drop in a photograph, say nothing and leave, and that is a complete record. Questions are offered and never gating. I track the proportion of one-line conversations, because if it ever reaches zero the floor has stopped being real.

The record has no prose field

An early version praised a child's progress on a fortnight of thin data, and it read as flattery, which cost more than saying nothing would have. The fix was structural rather than a change of tone. The generated record has no free-text field at all: it is assembled from coverage, typed observations and quotes copied out by code, and every entry carries a reference back to the thing it came from. Where her account and the worksheet disagree, both go in, unresolved. The coaching is written as prose, because it is a suggestion she can take or leave rather than a claim about her child.

Building it

I built it solo. Expo and React Native on the front, Supabase for data and auth, the Claude API for extraction and synthesis, deployed as a progressive web app so it installs on a phone without an app store. The hard part was never the interface, it was the schema. A worksheet page is not one piece of evidence, it is twelve, and any claim the tool makes months later has to be able to point back at the right one.

The habit that paid for itself was checking what the system actually wrote rather than what it reported. The day she submitted 34 pages, 10 of the submissions reported success and wrote nothing, from three separate causes, and fixing them recovered 230 lessons. A day later a refactor broke the live path while every test passed, because every test ran without writing, so now any change that writes to her record is run once for real before I call it done.

Constraints and trade-offs

Two things I decided not to build. The first is curriculum advice. Telling a tutor what a child ought to learn next makes the tool the authority, and the premise of the whole thing is that the tutor is.

The second is anything touching additional needs. I talked it through with a child educational psychologist I know and concluded that the distance between noticing a pattern and implying a diagnosis is shorter than it looks, and not a line for a tool at this stage to walk up to.

The pilot is one household, so the findings are deep rather than broad, and the household is mine, so her feedback came to me as her husband. Writing the accept criteria down before each session, and keeping the no when it came, is how I tried to stop that softening the evidence.

Where it stands

There are no commercial outcomes here. There is one household, four moderated sessions across ten weeks, and a pilot that is still running.

That first real conversation, on 21 September, is the most recent evidence and it cuts both ways. The coaching was grounded in her own records and she took most of it, but she also said she had no real interest in it creating a record, and that she would not use it again if she had downloaded it, because it had told her three times that it had saved lessons it never saved. Fixing that comes before asking her for anything more.

Before the current pilot window opened she recorded what she wants the app to produce. The window closes in mid-October, and I will judge it against that list rather than against my own sense of how it went.

What this shows about how I work

• I listen for where the value lands. I set out to build a long view of a child; she felt the value in everyday use. In July she wouldn’t pay for a tool like this; by September it was “the whole thing”. Felt value sits upstream of revenue, so I start there.
• Evidence over opinion. I wrote the accept criteria down before each session, so a disappointing result would be a finding rather than a matter of interpretation, and I kept the “no” when it came.
• Restraint as a design decision. The model can do more than the product lets it. I decided where it stops: it assists rather than teaches, it never goes near a diagnosis, and the record it produces has no free-text field to flatter with.
• I check what the system wrote, not what it reported. The day she submitted 34 pages, ten submissions reported success and wrote nothing. Fixing that recovered 230 lessons, and any change that writes to her record is now run once for real before I call it done.

Reflections

The rule that the record can only say what the evidence supports sounds like a values statement right up until you are the one writing the table that enforces it, at which point it becomes a column and a migration.

The other thing I had not expected to spend so much time on was the law. Holding information about a named child is legally fraught, and working through it turned what had been a design preference into the thing that makes the product defensible: Paideia does not assess the child, it supports the adult teaching them. It records what happened in a lesson, but every output is aimed at the tutor's practice rather than at a verdict about her child.

What is unresolved is whether any of it generalises. One household produced findings I trust and a sample I do not, and the only way to know which parts are about this family and which are about home education is to put it in front of households I am not part of.

The practices behind this →