RUMI
The users are curriculum teams, school-network leads and teacher trainers. They know their textbooks and classrooms well, but have never installed a developer tool or typed a command. They work across countries, subjects and languages, and many already have their own lesson-plan format. Science Fuse is the first trial partner.
The bet: if the pipeline sat behind a simple flow in the browser, anyone could bring their own book and leave with lesson plans they trust, without learning how the AI works underneath.
Seven machine stages became five human steps
The tool underneath runs seven stages: reading the pages, splitting the book into units, enriching, scripting, rendering, voice notes and delivery. Users never need to see them. I organised the flow around the only moments that need a human decision:
Add your book
Confirm the units
Set the timetable and level, and optionally add your own lesson-plan format
Review a sample lesson
Build the units you want, then download
Everything else runs in the background as the crew at work.
Two Interfaces, One Flow
GUI 1: the guided product. A calm interface in Rumi's brand. Folio, the host character, peeks over the upload box with eyes that follow the cursor and points the user to their first action. This prototype became the live product.
GUI 2: crew studios. The same flow, shown as a team at work. Each of the eight crew members has its own animated room with three states: working, idle, and "give me a minute". It tested one question: can watching an AI work feel like watching a team instead of a loading bar? Its ideas carried into the live product's progress panel.
Wireframe to live product in about 4 days, across 67 changes
Without AI, I estimate this would have taken at least three weeks, plus handoffs between a designer and engineers. More importantly, neither interface would have been approved or built this quickly. Ideas this ambitious usually get cut when priorities are set, or stay as pictures in a deck. With AI, both became fully working prototypes.
Prototype to product. GUI 1 became the app directly, with no spec and no handoff.
Design review in the real product. I reviewed live screens and changed them the same day. On 7 Oct alone I made more than 15 changes to the flow.
Characters and rooms. Crew Studios' eight animated rooms went from rough ideas to a full set in a couple of days, using AI image and video models.
Speed only helps if the judgment stays human. Three moments where I corrected or rejected what Claude Code produced:
The roaming character was cut (GUI 1)
Claude Code built a host character that wandered across the landing page. I cut it, because characters come after the screens, not before them. I kept only Folio peeking over the upload box, where it does a job: pointing the user to their first action.
"Lights off" for idle crew was rejected (GUI 2)
It dimmed the rooms of crew members who weren't working so there is emphasis on the crew member that is working and to give the user a sense of "moving forward". The page looked broken and empty. I replaced that with status labels, a glow behind the room that is working, and calm idle animations, so nothing ever looks frozen.
Fixes that only worked for one book were stopped (the flow)
When a sample book broke, it tuned its instructions to that one book. I set a rule instead: every fix must hold for any book, subject, language or lesson format. When a real book fails, we add a similar test book rather than tuning to the one that failed.
The crew asks when it isn't sure
If a book's structure is unclear, for example whether 1.0 is a unit and 1.1 a sub-unit, the user sees the crew's best guess as a question. A silent wrong guess would spoil every lesson after it, so the product shows its doubt instead of hiding it.
Nothing is built before the user sees a sample
The first sample lesson waits for Continue. Review is one decision, not two: the lesson sits on the left and the next step on the right. Users build only the units they need and come back for the rest.
Every stage became a named crew member (GUI 2)
The stages, plus a checker that validates the work, became eight characters: Folio, Tally, Verity, Quill, Inky, Hue, Echo and Dash. A long wait now reads as a team at work, and every member gets its turn in the live product's progress panel.
The first book is free, then it runs on the user's own AI
The first book is free. After that, it runs on whichever AI provider the user already pays for. Errors are explained in plain words, not error codes.
Accessible by default (GUI 1)
Every animation, including Folio's, respects the reduced-motion setting. Crew progress and Folio's chat are announced to screen readers. Buttons are labelled, and the main button sits in the same place on every step. Lesson plans can be produced in Urdu as well as English.
The product is live at curriculum.hellorumi.ai, and the first trial with real users, with Science Fuse, is starting now. It is too early for user outcomes. Tracking from upload to download, a rating on every lesson, and a metrics page for the team are already in place. The first numbers to watch are how many people get from upload to a downloaded lesson, and how reviewers rate their first sample.
The next design opportunity is the review step. That is where a curriculum lead decides whether to trust an AI-written lesson. With real reviewers in the trial, I want to learn what makes them approve a lesson, what makes them rewrite it, and what would make them trust the crew enough to build a whole book after seeing one sample.