Building with AI, a curriculum that school won't
School will give my son credentials. I'm building everything else.
School will give my son credentials. I’m building everything else.
By the time my three-year-old finishes school, the world will be unrecognisable. I don’t know what jobs will look like, or whether jobs in their current form will survive AI.
All education, not just college, is upstream of jobs: after the initial years, everything is optimised for curriculum meant to generate a certificate that is meant to get a job. Which means it’s also optimised for a version of the economy that may not exist by the time he graduates. School will prepare him for today’s world, which is all that it can do, in all honesty.
In fact by the time my son graduates, school may be solving the wrong problem.
It’s my job to prepare him for the future. But what does preparation looks like when the destination is unknown? I’m using Claude based orchestrator to help me create a parallel curriculum, reverse-engineered from the traits and skills needed to navigate a world no one can predict. While school will enable socialisation, peer friction, competition (hopefully), and the ability to navigate institutions, the content he learns in school will be searchable, or generatable.
The capacity to navigate uncertainty won’t be. I want to enable him to learn to learn.
The room that told me everything
I was surprisingly invited to a meeting for “SOAR: Integration of AI in Schools”, at India’s National Council for Vocational Education and Training (NCVET) last year. More than twenty minutes of the conversation was focused on whether they should rename a module from “ethical AI” to “responsible AI.” When it came to the delivery model, teacher training, implementation plan: “we’re working on it”.
Someone in the meeting pointed out they already use ChatGPT to prepare policy documents. “If you’re not, please start.” Then, in the same breath, said that children must still write assessments by pen and paper so we know they actually understood the material.
When I pointed out that explainability of AI, which is one of the things they were planning to teach, remains an unresolved problem globally, they acknowledged and moved on. They were focused on compliance. Someone at an AI Summit side event at the Canadian Embassy put it correctly: “Anybody who’s worked in the education system knows, it’s a big ship. It takes a wide berth to turn.”
I thus have to build my own. As I explained when I wrote about AI in higher education:
AI in education is most powerful for students who already demonstrate intent and curiosity, know how to think, question, and doubt, and it’s regressive for those who just want easy answers.
At a primary level, AI will empower willing individuals — parents and teachers — to enhance learning for Children.
What does a parallel curriculum look like?
A friend who intermittently home-schooled his kid, told me that they turned the walls of their house into whiteboards. Dinner table conversations included function-guessing games: give me an input, I’ll give you the output, figure out the rule. His daughter was doing algebra at six. He’s homeschooling her now. We don’t want to homeschool, but still, my job is to focus on skills through projects, play and discovery.
Using Claude, I’ve built a 30-skill orchestration system that guides me in guiding him. The 28 skills break across five categories:
Thinking skills: First principles thinking, second-order effects, systems thinking, reasoning (inductive, deductive, probabilistic), critical thinking, innovation, among others.
Social-emotional and life skills (9): Including emotional regulation, entrepreneurship fundamentals, leadership, Public Speaking. Recently I’ve added “comfort with ambiguity” to this list. These are areas where schools can’t do much because they’re hard to measure and certify.
The remaining Claude Skills help with implementation:
Assessment skills: Documentation analysis, mastery assessment, struggle detection, gap identification, learning velocity tracking.
Planning skills: Monthly review and planning, materials recommendations, life skills workshops, among others. The output of each monthly cycle.
Projection skills: Future trajectory mapping, and how to succeed inside a compliance system without being defined by it.
In the system, for privacy, my kid is cheekily named Vikas. Through the month, my wife and I document observations, sometimes from Parent-Teacher Meetings, which I input into Claude: what happened, what worked, what didn’t, new behaviours, struggles, interests.
Claude analyses using all 28 skills, identifying patterns we don’t know about. We receive the full analysis, which includes:
- Granular breakdown of development by domain (and which are realistically plausible at his age. Many are not yet)
- Realistic advancement assessment relative to developmental expectations
- Active developmental windows (critical periods where specific focus yields disproportionate returns)
- Trajectory progress toward age 5 and age 9 goals (which we have set)
- Risk assessment with mitigation strategies
- Next month’s activity schedule, often flexible but specific, along with things we can buy, and what not to buy.
Through the month, I seeking advice for activities, including those not in that schedule.
A critical feature is correctability. When I tell the system “Vikas” doesn’t count to ten reliably (he adds 11, his favourite number, quite randomly), it updates the tracker, revises assessment, and adjusts projections, and offers activities that we can focus on. Accuracy of data improves over time, but we’re focused on small steps, compounding his learning, and mastery. A wonderful year at Learning Matters, which operates on the Reggio Emilia philosophy, has given him a sense of agency, which we want to encourage.
He builds remarkably complex structures with Magna Tiles, takes a complete-destroy-start again approach. Nothing persists. I’d like him to work on long term projects by the time he’s nine, so the shift from instant gratification to sustained building is important. Since he loves Aeroplanes, the orchestrator suggested we build an airport, but one component per week. We have three so far - a runway, a hangar and an air traffic control tower. We will graduate to LEGO’s soon. LEGOs check two of three boxes that I’m focused on: Play, projects and discovery.
Because he is curious about machines, Claude suggested we introduce him to the “How Things Work” books, which are typically for 8-9 year olds. He can’t read so I explain them to him.
AI will amplify the ability of parents who know how to design systems.
The models I’m learning from
The best learning systems don’t look like schools. As a teacher as Learning Matters said today: focus on play and wonder. The question isn’t about just what to teach, but also how learning is designed.
At an AI Summit session by LEGO that I attended, an executive from the LEGO Education Foundation pointed out that:
“AI cannot teach curiosity. AI cannot teach empathy. AI cannot teach creativity. What we can do is create environments where children experiment, fail, collaborate, build, and question.”
AI systems today are optimised for frictionless completion: reduced time-to-answer, higher engagement, faster resolution. Dopamine hits, like Social Media algorithms. Development requires the opposite: friction, struggle, boredom, social negotiation, the experience of being wrong in front of people you care about.
We need to build grit, the tolerance for struggle, the joy and beauty of building through that struggle over time, sometimes years. I know this as an entrepreneur of 18 years: constant experimentation, optimisation, failure, building new systems, dismantling old ones, learning all the time.
A large part of what I’m building for myself, imperfectly draws from Alpha School, a mastery based school in Austin, Texas. They have academic learning for two hours a day. Third-party MAP Growth tests show Alpha students achieving 2.3x annual growth compared with peers, completing a grade level’s worth of progress in roughly 22 hours of focused study.
The mechanism:
Remove the pacing constraint of the median.
Add immediate feedback loops.
Advancement of levels is based on genuine mastery rather than time-spent.
So it’s personalised, and not like the traditional mechanism of moving cohorts forward by age. Alpha School’s the mastery approach: replace “I’m just bad at math” with “I haven’t learned that yet.” Very Mindset of them.
The rest of the day, Alpha School students spend on life skills workshops: entrepreneurship, public speaking, leadership, real projects with stakes. Learning, not compliance.
We’ve chosen compliance for “Vikas”, and he starts in Nursery this year, but I’m trying to learn as much of this model as I can, so that we can focus on mastery and the “beyond academics” piece at home.
What about AI for Learning?
I’ve seen kids who can’t eat their meals without YouTube open, which is why we’re currently on a “no-devices, no-sugar policy” for now. At the same time, discovery is limited by lack of devices. Much of what I’ve learned has came from exploring the internet as an autodidact. AI can be even more empowering for builders.
The same tools that create dopamine dependency can, if designed differently, improve learning outcomes. That is why I want to also eventually enable AI based learning for him, like Alpha School, Khan Academy and Math Academy do: focus on improving learning outcomes without the dopamine hits.
We’ll take it slow: first introduce devices, then learning through engagement and experimentation through devices. I can’t wait to get to the Aurdino kits. I’ve only just started playing with the Raspberry Pi myself.
What I’m doing here is probably early, but not for long. There’s will be a demand for systems like this: flexible, adaptable, and based on learning outcomes not taught in schools. Built around the constraints of a parent.
The problem is that it requires parents with the privilege of time, knowledge, and disposition to engage consistently with the child.
AI doesn’t replace parenting or teaching. It scales intentional parenting and teaching.
I’d love some feedback and inputs: what do you think about what I’m doing. What do you think I should be doing differently? What are you doing that I can learn from? Leave a comment or drop me an email at nikhil at medianama dot com.