AI owns
- Lesson interpretation
- Vocabulary selection
- Contextual translation
- Visual concept planning
- One illustration per concept
VOCABULARY STUDIO · PRODUCT NOTELive working prototype
Designing a multilingual lesson-adaptation feature for a CoTeach-style workflow.
How I moved from an initial AI prototype to a more credible classroom product—through quality failure, model iteration, cost judgement and a deterministic worksheet workflow.
Standalone concept designed to fit a workflow like CoTeach. Not integrated with or endorsed by CoTeach.
01 · INITIAL PRODUCT HYPOTHESIS
The useful AI workflow was not “generate a pretty worksheet.” It was understanding the language barriers inside an existing lesson.
I wanted to showcase a focused feature that could fit naturally after lesson creation. A teacher should be able to take the lesson they already have and create a multilingual visual vocabulary resource without re-entering every word.
That led to a deliberate separation: AI interprets the lesson and creates structured educational content; the application protects the worksheet structure.
The application owns
- Fixed worksheet layout
- Accurate editable text
- Progress and loading states
- Retries and partial failure
- PDF and PowerPoint export
02 · WHAT FAILED IN THE FIRST ITERATION1 · Workflow fit 2 · First build 3 · Quality failure 4 · Product decision
The first image route returned technically valid pictures that did not reliably communicate the vocabulary meaning.
Place the feature after lesson creation, as a downstream adaptation layer.
Selection, translation, independent images and structured exports worked end to end.
The pictures looked acceptable but did not reliably communicate the vocabulary meaning.
I rejected the route despite its low request cost. Technical completion was not pupil value.
“Would this image help a multilingual Grade 1 pupil understand the word in this lesson?”
03 · MODEL AND COST ITERATION
Optimise for cost per usable classroom asset—not cost per request.
I used the smallest model route that met the relevant quality threshold in live testing. Both choices remain replaceable environment variables rather than fixed product assumptions.
Flux 2 Klein 4B → Gemini 3.1 Flash Lite Image
The first route was cheap and technically valid, but relevance was unreliable. Gemini produced a lesson-relevant “add” visual showing two groups joining.
Claude Sonnet 4 → Gemini 2.5 Flash Lite
The task was constrained structured extraction and translation. The smaller route met the live benchmark at a much lower token cost.
04 · WHAT THE FINISHED PROTOTYPE DEMONSTRATES
A lesson-grounded workflow with teacher control, bounded generation and recoverable failure.
Contextual language
Reads the complete lesson and selects up to eight learner-critical terms.
Bounded concurrency
Up to four image requests run at once and cards populate progressively.
Recoverable failure
One failed illustration can be retried without discarding seven successful cards.
Structured output
PDF is print-ready; PowerPoint keeps text, shapes and images editable.
Review and guardrails
Optional vocabulary review and age-appropriate image constraints.
Cost controls
Replaceable models, request limits and a generation kill switch.
05 · WHAT I WOULD BUILD NEXT