One Coach per Subject: Why 30+ Specialists Beat One Giant AI
A persona with one job can hold standards a do-everything assistant cannot. Inside the design decision that shapes everything else about OpenKids AI.
A Persona With One Job Can Hold a Line
OpenKids AI runs on one coach per subject — 30+ specialists across 6 academies — instead of one giant AI that does everything. The reason is not aesthetic. A persona with exactly one job can hold standards that a generalist, by construction, cannot.
Our maths coach never writes your English essay, no matter how the request is phrased, because essay-writing sits entirely outside its role — there is nothing to negotiate with. Our Chinese coach stays in Chinese even when a child tries to drag the conversation into English, because staying in the language is the pedagogy. A do-everything assistant has to decide, case by case, whether to comply. A specialist never faces the question.
Depth You Can Actually Train
The second advantage of narrowness is depth. Because each coach has one subject, its teaching style can be written, reviewed, and refined the way a head of department coaches a young teacher. How the maths coach introduces bar models before algebra, how the writing coach responds to a bland opening sentence, how the science coach handles a why-chain from a curious ten-year-old — these are authored decisions, not model improvisation.
Spread that same effort across one persona that teaches everything and it thins into wallpaper. Concentrate it, one subject at a time, and you get something closer to craft. Depth of persona is the real moat here: a teaching style you can deliberately train beats raw capability that arrives with no opinions about teaching at all.
Specialist Tools, Not Generic Chat
Real teachers carry subject-specific equipment, and so do our coaches. The maths room comes with interactive widgets built for how mathematical understanding actually forms: fraction bars a child can drag until three-quarters stops being a symbol and starts being a size, an equation balance that makes both sides of the equals sign feel like physics, a times-table grid that turns memorisation into pattern-spotting.
A generalist chat window cannot justify building any of this, because no single tool matters to more than a sliver of its conversations. When the coach serves one subject, the economics reverse: every hour invested in a fraction bar pays off in every fraction lesson, forever.
Failure Modes, Anticipated per Subject
Every subject fails in its own way. Children misread fraction addition as adding tops and bottoms; they treat the equals sign as ‘the answer goes here’; they memorise 化学 characters by shape and collapse at 听写. A generalist meets each of these cold, every time.
Our coaches are built with their subject's known failure modes written in: the misconceptions to probe for, the seductive wrong methods, the step where this topic usually breaks. When a P5 student's answer is wrong in a classic way, the coach recognises the classic, not just the wrong. Anticipating the specific mistake is most of what expert teaching is.
Real Expertise Is Narrow
No school hires one teacher for every subject, and not because schools lack imagination. Centuries of running classrooms taught them that real expertise is narrow: command of a subject's sequence, its tools, and its characteristic confusions does not transfer just because the person is generally clever.
That is why the school metaphor is not branding for us; it is the architecture. One coach per subject is simply what taking teaching seriously looks like when the teachers are made of software. The generalist AI is a brilliant polymath with no lesson plan. We chose the faculty.