Any teacher who has photocopied thirty identical worksheets knows what happens next. One child needs the instruction shown rather than described. Another writes perfectly well but cannot hold to the line, and by the fourth question the page has gone crooked and the willingness has gone with it. A third stops early, not because the questions are hard but because there are twenty-eight of them crowded onto one sheet. Teachers see all of this, every day. What they have rarely had is a practical way to act on it thirty times over before Monday morning.
So we built the other half of it. Where a school has already recorded what its class teachers have observed about a child, the worksheet builder can take that record and produce a sheet for that child alone — the same subject and the same chapter as everyone else's, at the grade the school has recorded for that child, worded and laid out for the person who has to sit down and do it. It covers grades 1 to 5 in every school, and grades 6 to 10 where the school has been switched on for them.
It starts with what the school has already written down
The builder works from what a school has recorded about a child rather than from any judgement of its own: how a child learns best, where they are working academically, what interests them, and the specific things a class teacher has written down about how they handle written work. A school can hand over the spreadsheet it already maintains, as an Excel workbook, or type the same handful of fields in; both routes end at exactly the same validated profile, so the two surfaces cannot disagree about what a complete one looks like.
Two habits of a real school record turned out to matter. The first is that a child usually has several things noted rather than one, and a school's sheet often spreads them across several columns — so every column that reads as an observation is collected, in the order the school itself ranked them, instead of only the first. Reading one column addressed at most one recorded need per worksheet and quietly dropped the rest, which was the single largest source of adjustments that never reached the paper.
The second is that many schools already keep a column of recommended adaptations. That column is captured, but it is deliberately not what drives the sheet. The observation is the source, and the response to it is looked up separately, in a shared reference table of what to do about each one. That way a school can be shown where its own file and that table disagree, rather than having a hand-typed phrase silently decide the shape of a child's page.
One sheet per child, not one sheet annotated afterwards
Each child's worksheet is built on its own. The curriculum material for the class's topic is retrieved once for the whole group, so every child is working from the same chapter of the same book; what differs is how the questions are put and how the page is laid out. Each sheet is generated, checked and saved independently, so one child's sheet failing to build leaves the rest of the class unaffected.
Before anything is saved, the generated questions are inspected for a specific failure: an item that cannot be attempted at all. A colouring task with nothing to colour, a dot-to-dot with no dots, a question with an empty prompt, a matching pair whose two sides are the same word — each is well-formed and each prints as a numbered blank. Anything caught is rebuilt once, with the format that failed taken off the menu where the failure was the format itself, and whatever a rebuild cannot rescue is removed rather than handed to a child as a blank they are expected to fill.
What actually changes on the page
That reference table — written out of a real school's own workbook — pairs each recorded observation with a strategy. Much of what those strategies ask for is not something you can say to a generator at all: it is ink a page has to lay down. Ask a language model for a shaded margin zone and you get the words "shaded margin zone" printed on a child's worksheet. So the page draws most of them itself:
- A worked example at the top, solved step by step, before the child's own questions begin
- Instructions written as short checklist cues rather than a paragraph of prose
- One question to a bordered box, one box to a row, instead of a dense column
- A short panel at the top explaining how to do the sheet, and a bank of the target words above the first section
- A progress strip — so many done of so many — with a tick beside each finished question
- A marked micro-break after every few questions
- A line on each question for the child to note that they checked it, and a review step at the end of a section
- One square per letter instead of an open line, so a skipped letter leaves a visible gap rather than an invisible one
- Wider ruled lines, a tinted band down the left margin, and a dot between words wherever the sheet asks for writing
- Larger type, more space between the lines, and a wider margin to write into
- Pure black on white — no grey rules, no tints, and no instruction that depends on telling two colours apart
- An empty framed area to draw the answer in, for a child who thinks with a pencil in hand
- Questions set in the things the child actually likes
The same work, not easier work
This is the part we got wrong first and now care about most. A great many of these observations are about how a child handles the page — grip, movement, repetition, letter formation — and not at all about what they can read or reason about. Left unsaid, an instruction to include tracing and counting activities reads to a generator as an instruction to drop the content to an earlier year, and that is precisely what came back: children reading comfortably at their own grade were handed a sheet asking them to trace a letter.
The builder now carries an explicit counter-instruction naming the child's own grade. Keep every item at full grade level; change only the format; do not simplify the vocabulary, shrink the numbers, or fall back to earlier-year topics. The same principle governs the activity types: a recorded preference for hands-on work now adds hands-on activities to what is available, and never removes the grade-level ones. It used to remove them, and a sheet built that way had no way to obey an instruction to stay at grade level even while being told to.
There is exactly one entry in the whole table that genuinely licenses easier content, and it is the school's own recorded judgement that a child is working below grade level and needs structured extra practice. That is a teacher's professional call, made before any of this runs. It is not something the software infers from the fact that a child finds it hard to sit still.
The sheet never says why
Nothing on a child's page names a difficulty or the adjustment made for it. The only information about the child that may be printed is their name, the school's own student reference, and their grade. The observation, the strategy and the wording of the recommendation are all stripped out of the generated text before the sheet is saved, and the check runs in both directions, so a fragment or a truncated echo of one of those phrases is caught too. The child's name is not sent to the model either — it is attached to the sheet after the questions come back.
The reason is not technical. A worksheet is handed to a child in a classroom in front of other children, and a printed disclosure cannot be withdrawn. Where the school's administrator needs to know why a sheet looks the way it does, that answer lives in an admin-only view, alongside an honest list of which entries actually fired for that child and which recorded phrases matched nothing in the table and need a person to look at them. A phrase that matches nothing is reported rather than swallowed, because a stale or renamed wording is otherwise an accommodation a teacher deliberately recorded and silently lost.
The child's interests are the one deliberate exception to all that stripping. A worksheet built around cricket is the whole point of having asked.
On paper and on screen
The adjustments travel with the sheet rather than with the place it happens to be displayed. The printed page and the on-screen version a child answers in are drawn by two different renderers working from one shared list of adjustments, so an adjustment is not simply forgotten when the same sheet is opened on a tablet. Where an adjustment cannot honestly mean anything on a screen — a frame to draw in, when there is nothing in hand to draw with — the on-screen version says so plainly instead of showing an empty box that does nothing. A few belong to paper by their nature: ruled lines, margin bands and type sizes are properties of a printed sheet, and they follow it wherever it is printed from, including the copy a student prints for themselves.
A draft, and a person
Personalised sheets are built from the school's own administrator account — not by teachers, not by department heads, and not by parents or students. A list of children with their recorded difficulties attached is the most sensitive roster this platform handles, and it stays with the account a school already holds accountable for its records.
Every sheet lands as a draft, in the ordinary worksheet library, assigned to nobody. It can be previewed, edited question by question, regenerated, printed or thrown away, and it reaches a child only when someone deliberately assigns it. That ordering matters more than any single adjustment above: the software proposes a page, and a person decides whether it is the right one.
That review has already changed the feature. When teachers went through the first batch of real sheets item by item and told us which questions were unusable and which of a child's recorded needs had never actually reached the paper, several of the adjustments described here did not yet exist. A setting being switched on is not the same thing as ink on a page, and only a teacher reading a printed sheet can tell you the difference.
None of this measures a child or places them on a scale. There is no score here and nothing that ranks one pupil against another. There is a set of things a teacher noticed, a table of what to do about each of them, and a page — asking the same thing of the child as everyone else's page, on paper they can actually work on.