How Goodnotes did it the slow way — building with teachers — to find the real need behind AI marking and a handwriting tutor.
Human methods, new material: AI for the classroom ft. Mubarak Marafa & Wendy Liao | Config 2026 · Mubarak Marafa
21 min total·Actually worth watching closely: ~7 min·3 must-watch clips
- 0:14 – 3:00Listen
An unexpected starting point: 3.5 million teachers
Of 27 million monthly active users, about 3.5 million turned out to be teachers, using Goodnotes in surprising ways nobody expected — which is what led to building Classroom. And the claim that frames the whole talk: great products are not built for people, they are built with people.
The real risk in the age of AI is letting your humans get too far away — go and find your users first.
The opening is spoken narrative with no key visuals; fine to listen to while you get oriented.▶ Jump to 0:14Speaker · Mubarak Marafa - 3:00 – 7:00Listen
Into the classroom: the truth interviews won't give you
In standard interviews teachers tell you what they think they do. Only standing side by side with them surfaced the real workflows — unstapling assignments and re-sorting them into piles to mark in context, reaching for a tool for the analysis it produces rather than the tool itself — along with the full before-work, during-work, after-work journey.
A surface-level request pulls an AI product off course: what teachers want isn't the tool, it's the analysis behind it.
This stretch is pure spoken methodology with no visual dependency — dense with ideas, so listening pays off more than watching.▶ Jump to 3:00Speaker · Mubarak Marafa - 7:00 – 9:20Watch
The first features: swipe marking and split view
Classroom observation turned straight into product: one gesture swipes a teacher through an entire classroom of work, and built-in screen share displays it while protecting student privacy; the split view escapes the physical limitations of real-world paper so teachers can mark a whole class in context in a single view.
The split view became the cornerstone of the product, giving every student a more equal shot at instant response from the teacher.
Three visual moments land here (the swipe gesture, the screen share, the split view) — this is insight becoming interface, so watch the screen.▶ Jump to 7:00 - 9:20 – 12:00Watch
AI enters: clustering handwritten answers for marking
Handwriting recognition understands the question and the students' handwritten responses and automatically sorts them into clusters; teachers mark the clusters and it's applied to all their students, with instant results on which students or which group of students got the correct answer.
The right entry point for AI is killing the repetition in marking, not making the teacher's judgment for them.
At 632s the full clustering-marking flow is on screen; "mark once, applies to the whole class" is hard to feel from description alone — worth watching the demo.▶ Jump to 9:20Speaker · Wendy Liao - 12:00 – 14:15Listen
The secret sauce: group sessions beat one-on-ones
Interviewed alone, teachers give idealized, neat versions of their workflows, or leave out the little hacks they've internalized as just part of the job. Put them in a room together with coffee and they set each other off, and the real workflows come out.
To get at real needs, let users talk to each other rather than questioning them one at a time.
The research methodology is mostly spoken with nothing to demo — fine to treat like a podcast.▶ Jump to 12:00Speaker · Wendy Liao - 14:15 – 18:20Listen
From users to partners: showing up and co-creating
The team kept going into classrooms and saw first-hand what a bug really costs — an app crash mid-lesson and 30 ten-year-olds lose focus in a second, the lesson stops. Then 40 teachers were invited into the office for a jam session, and teachers went from adapting to the app to helping shape it.
Only by being there do you feel what a bug costs in a real classroom — data from a distance will never show you that.
Mostly story and argument with no visual dependency; the narrative itself is where the information lives.▶ Jump to 14:15Speaker · Mubarak Marafa - 18:20 – 20:45Watch
AI accelerates create: from gaslighting ChatGPT to a handwriting AI tutor
First, prompts to gaslight ChatGPT into being a teacher and explore the approach; then a vibe-coded rough prototype sent to a real user to iterate on. What shipped is the AI tutor — teachers imbue intelligence into a question, and the AI sees what the student is writing and gently guides them toward the answer instead of giving it.
What AI accelerates is the create stage: the prototype doesn't need to be pretty — get it to a user, validate before you polish.
Three visual moments cluster here (1108/1151/1174s); the vibe-coded prototype and the AI tutor on screen explain the design trade-offs better than the words do.▶ Jump to 18:20Speaker · Mubarak Marafa