Teacher development is one of the few professional disciplines where the feedback loop is measured in terms rather than days. A teacher may be formally observed once or twice a year, by one colleague, for one lesson — and the notes arrive long after the class they describe.

Meanwhile the tooling that has arrived in classrooms tends to solve a single slice: a lesson-plan generator here, a transcription tool there, a quiz builder somewhere else. Each is useful on its own and none of them talk to each other, so nothing accumulates. The teacher does the integration work by hand, or the tools go unused.

I work as the Business Analyst and AI Product Designer on this platform: the person sitting between what teachers actually need and what the models can reliably produce.

That means the ambiguity arrives with me. I take a broad ambition — make teachers better, using AI — and turn it into module boundaries, data contracts, prompt behaviour, acceptance criteria and screens an engineering team can build against without guessing.

The framing question was not "where can AI help a teacher?" but "which part of the teaching cycle currently has no feedback at all?" That reframing is what moved the product away from a lesson-plan generator and towards rehearsal and evaluation, where the gap is real.

Decomposing a lesson into its observable behaviours — talk ratio, question density, wait time, the order an explanation is built in — gave the platform something measurable to work with. Those became the evidence layer that every downstream judgement has to cite.

The second decision was to treat the existing deck as the only required input. Any product that asks a teacher to author something new before they get value will lose to the Sunday-evening version of that teacher.

A closed loop of connected modules rather than a chat assistant. The teacher uploads the material they were already going to teach; the system reads it the way a mentor would, builds a classroom around it, listens to the rehearsal, then scores what happened against what was planned.

The important design property is that each stage produces structured output the next stage consumes. Lesson analysis is not a summary for a human to read — it is a map the classroom simulation is built from. Evaluation is not a grade — it is bound to transcript spans so a teacher can see the exact moment being described.

The platform is specified module by module, each with its own PRD section, data contract and acceptance criteria. Keeping the modules separately specifiable is deliberate: the classroom simulation and the evaluation engine have very different reliability characteristics and should be able to change independently.

AI behaviour is treated as product surface rather than configuration. Prompt design, output schemas and guardrails are specified with the same care as screens, because for this product they are what the user actually experiences.

Testing logic focuses on whether generated output is usable in a real classroom rather than whether it reads well. A lesson analysis that sounds articulate but produces no teachable decision is a failure, even though it would pass a casual review.

The recurring iteration is towards specificity: early outputs described teaching in general terms, and each pass has pushed the system to cite the actual moment, name the actual misconception, and propose something a teacher could do on Monday.

The hardest requirement work was not eliciting features but resisting them. Almost every stakeholder conversation surfaced a plausible new AI capability, and most of them would have made the loop longer without making it better.

Designing AI behaviour turned out to be closer to writing acceptance criteria than to writing prompts. The useful question is never "does this output look good?" but "what decision does a teacher make differently after reading it?" — and that question kills a lot of impressive-looking output.

Framing matters commercially as well as technically. Described as an AI chatbot for teachers, this product competes with everything. Described as a connected system covering before, during, after and over time, it occupies a space most tools do not attempt.