Concepts
Soon, this will all seem obvious.
Six ideas behind how we build stylus. Each took a lot of work to reach, and each will seem obvious once you’ve seen it. Dom calls that the second kind of obvious.
Scroll to begin ↓
01
Service, not a tool
People will soon say
“Remember when we bought a tool, and then did all the work ourselves?”
Service, not a tool.
We’re an education company, not a technology company. We use technology to make the education work.
In the world
You don’t buy the helpdesk. You buy the answer.
Customer service used to mean buying software, training your staff in it, and having them do the work anyway.
Intercom’s AI agent, Fin, charges per resolved customer query. Not hundreds a month for a tool: dollars per happy customer.
Buying a tool
- Seats and licences
- Training and set-up
- Your people still do the work
- You pay whether it works or not
Buying an outcome
FinIntercom’s AI agentNot resolved? Not paid.
In schools
Pay for the marking, not the marking software.
Teachers scan the work in, and that’s it. We do the marking, the moderation, the analysis and the quality assurance, then hand back educational decisions, not software to operate.
Nothing turns up in front of a teacher unless someone has looked at it.
The tool way
The service way
“So it’s a tool?”
No. A tool makes the teacher the arbiter of what’s good, and it may as well be ChatGPT with their pupils’ names in it.
Software is cheap now. The work that makes it good is still very much in the human domain, so we take responsibility for the quality of the service.
02
The humble piece of paper
People will soon say
“We needed laptops, logins and training just to find out how a child writes?”
It starts with the humble piece of paper.
We’re the paper people.
In schools
Print it. Hand it out. Write. Photocopier.
We all start with the lowly, humble piece of paper. No new workflows, no training and no screens, for teachers or for pupils.
The lowest common denominator matters: even the most reluctant teacher can put a class set through the photocopier, and the trust still gets the insight. Our accessible papers speak to children where they are, rather than where we expect them to be.
03
Capture the detail
People will soon say
“We only wrote down the summary? We threw everything else away?”
Capture the detail, not the grade.
It’s easy to create a grade from the detail. You can never go the other way.
In the world
Keep the conversation, not just the doctor’s note.
A doctor’s note is three lines, filtered through one person’s attention at the end of a long day.
We settled for summaries because they were all one person could write down, not because they were better. Keep the conversation and the summary is still there, along with everything the note left out.
The note
Three lines, at the end of a long day.
The consultation
Every sentence is assessed against every one of our 127 criteria: evidence, error, or not applicable. That is tens of thousands of data points for a single piece of writing.
Meanwhile a Year 6 teacher can spend more than 200 hours on a writing portfolio to arrive at “uses the range of punctuation mostly correctly”. What does “mostly correctly” mean?
One piece of writing
“Fundamentally, Dominic, nothing matters until they’re writing in complete sentences.”
Dr Tim Mills · so we count complete sentences, not grades
Deceptively simple
Then the hard part: deciding what matters.
Our reports are built from the ground up. First, everything that could be said. Then the one thing that matters most on this child’s learning journey. Only then does an AI write the comment.
Ask an AI to comment on a whole piece and it is just imitating educational thought. We’ve done the educational thought.
Everything that could be said
The one thing that matters most
Full stops between sentences
Because nothing else lands until it does.
Only then, the comment
Read your second paragraph aloud. Where you pause for breath, does a sentence end? Put a full stop there, and start the next with a capital.
04
Formative first
People will soon say
“We judged progress by grades alone, and left teachers to guess what came next?”
Formative first.
Find the thing that’s going to unlock them.
In the world
Tell it where you’re going. It routes around whatever’s in the way.
A sat nav doesn’t just report where you are. It keeps adjusting to what’s ahead, and reroutes before you reach the queue.
A blocker in a child’s writing is that queue. Route around it early, or accept that they arrive late.
In schools
Where they could be, not just where they are.
Look at what a pupil can do across a number of pieces of writing, and find the blocker. Then tomorrow, a twenty-minute carousel: one group working independently, one with a TA, one being retaught, “because of this that we’ve seen in their writing”.
That was never feasible for one teacher before.
Summative first
Formative first
“I’m working on my adverbial phrases.”
A boy in Tamworth, pulling his stylus report from the back of his exercise book
05
Expert evaluations
People will soon say
“We just trusted whatever the model said? Nobody checked whether the experts agreed?”
Expert evaluations.
Quality is everything. Test the model against the people who know, one decision at a time.
In the world
The AI you ask about your health is now marked by doctors.
For OpenAI’s HealthBench, 262 doctors who have practised in 60 countries wrote down what a good answer looks like across 5,000 realistic health conversations: 48,562 specific criteria.
The question is not whether the model sounds confident. It is whether the doctors would agree.
A model’s answer, waiting to be marked
Written by doctors
Source: HealthBench, OpenAI, May 2025.
In schools
We’ve gone slow to go fast.
We marked everything by hand first. At GCSE that meant a month of full human marking, then moderation, then randomised moderation. We spent two back-to-back DfE contracts getting the marking good enough before we sold it.
Every disagreement between the model and our expert moderators is reviewed and written back into the guidance.
One decision at a time
Year 10 · descriptive writing
…past the old, corrugated, iron, shed by the gate…
The model
Expert moderators
Disagreement reviewed
“Corrugated iron” is one compound adjective: hyphenate it, and no comma before the noun. Guidance updated.
Agreement with the consensus of our expert moderators
Same sentences, same criteria. More consistent than any one marker, and every disagreement goes back into the guidance.
The bar is higher than it looks: two experienced examiners marking the same English script correlate at only around 0.4.
06
Outsource the work. Never the thinking.
The war on AI slop, first declared at MATPN South, 2025.
Our position is that we should all be working to elevate and support the role of the teacher, rather than circumventing it.
People like Dom left the classroom because they spent most of their time holding a red pen. Take the red pen away and teachers can read the writing with a cup of tea and enjoy it, because the AI will never enjoy the writing.
That’s why we’re at war with AI slop: unchecked model output put in front of children, or feedback one trust leader called “soulless”. What every child needs isn’t an AI tutor. It’s an inspirational teacher with the clearest possible picture of their writing. The reasoning behind every sentence works as professional development: we upskill teachers, we don’t deskill them.
Our rule for anything we make
It should take longer to make than it takes you to read.
Put together
Each idea makes the next one possible.
Less time measuring pupils. More time improving them.
It’s on us to prove it in your context.
Start with a year group, or a few classes. Or read who we are, and why we work this way.
