What does it suggest? We iterated on default question chips — pre-built prompts that gave teachers a starting point instead of a blank text box, which our data and early interviews both pointed to as something teachers actually used to get oriented.

When does it show up? We explored several just-in-time placements across the curriculum, but ultimately decided to keep the first release at the Lesson page. This was a direct, in-the-moment call made jointly with engineering rather than a rule handed down from a spec: too broad, and it competes for attention against everything else on the page; too narrow, and teachers never discover it.

How far do we push visual identity? Partway through, there was a real pull toward a fuller redesign — more distinct branding, more of an "agentic" personality. We paused and made the call to scale that down to targeted refinements instead of a redesign this close to launch, and communicated that explicitly to the team: the underlying interaction was sound, and a redesign right before shipping would have traded a working thing for a better-looking maybe.

Results Since Launch

We tracked adoption and feature retention weekly from GA onward.

Adoption grew steadily and nearly doubled the original 15% target within about five months. Feature retention held in a fairly narrow band the whole way — it dipped slightly as adoption broadened past the earliest, most motivated users, which reads as normal dilution rather than a quality problem. The clearest signal in the data: teachers from the original early-access research cohort retained at 57–60%, more than double the general population's ~24–25%, pointing to an onboarding or familiarity gap rather than a flaw in the assistant itself — a good candidate for the guided setup flow I'm working on now. The most recent trend worth watching is rising latency (up from 2.0s to as high as 7.0s), which the team has flagged as the active risk to protect the helpfulness and retention gains that have otherwise held up well.

Early Access Research

Once real teachers were using it, I ran a round of 6 virtual 1:1 interviews and a school visit during the early-access window.

A few findings reshaped how we thought about the whole product:

Thread abandonment was high. A large share of conversations were started and never finished. That's an uncomfortable number to sit with, but it was more useful than a vanity metric would have been — it told us teachers were finding the entry point but not consistently finding a reason to stay.

Answer quality wasn't the bottleneck — output format was. Teachers liked what the assistant told them. What they actually wanted was something they could immediately use: a printable worksheet, a modified slide, a set of practice problems — not another paragraph to transcribe into their own materials by hand.

We were competing with a workaround teachers had already built. Some of our most AI-comfortable teachers were already downloading our curriculum content and feeding it into ChatGPT or Gemini themselves to get exactly this kind of help. That reframed the goal: this wasn't a feature competing against "no AI," it was a feature competing against a workflow teachers had already assembled on their own.

One assistant, several very different jobs. Veteran teachers wanted fast artifacts and summaries. Instructional coaches wanted unit-level strategy and standards alignment. Newer teachers wanted scaffolding and reassurance. Designing one interface to serve all three without picking a side became the next real problem.

Trust is a dependency, not a detail. Teachers were candid that typos and inconsistencies elsewhere in the platform made them hesitant to trust an AI feature layered on top of it. That was outside the scope of this project to fix directly, but it mattered for sequencing — an AI feature can only be as trusted as the product it's built on.

“Sometimes we’re jumping in and out of different units… so it becomes limiting when you can only look at one at a time.”  

Turning Findings Into a Roadmap

Rather than hand engineering an unordered wishlist, I worked with the team to structure what we'd learned into a three-phase direction:

1. Artifact Engine — close the format gap first: printable outputs, slide remixing, multi-lesson practice generation.

2. Memory Layer — give the assistant a sense of continuity: what a class already covered, what to review, support for teachers who don't teach lessons in a straight line.

3. Curriculum Strategist — the long-term version: standards-aware guidance and data-informed recommendations at the unit and course level.

Where we are now

Working with the go-to-market team, Kiddom Assistant is one of the most useful and popular features Kiddom has.

We are experimenting with a more fullscreen experience to make the output and artifacts easier to read. We will learn from this iteration and see