The workflow before the product.
Creating a curriculum-aligned exercise required both content writing and pedagogical judgment. The opportunity was to help the team produce a useful draft more quickly while retaining a deliberate quality check.
An early prototype also exposed a practical problem: repeated questions and themes. I discussed adding thematic context with engineering while working out how the generated content would enter the learning workflow.
What I owned.
I drafted the AI squad’s initial framework and feature options, then worked through the direction with our head of product. I led prompt iteration, coordinated pedagogy review and worked with engineering to deliver the authoring workflow.
I worked alongside the engineer who implemented and deployed the initial generation engine. My part covered the product direction, prompts, expert review and delivery coordination.
The decisions that mattered.
A strong instruction does not rescue a weak passage.
Pedagogy feedback showed that instructions could work well while the passage was too direct to support meaningful interpretation. I proposed reviewing reusable instructions separately from generated passages and questions, then revised the passage prompt for another expert review.
Keep prompt changes comparable.
I asked the team to retain earlier prompt versions so changes could be compared. We worked through successive sample batches with pedagogy feedback. When another generated passage might still fall short, I raised curated source material as a possible fallback.
Make the handover part of delivery.
I coordinated walkthroughs with engineering and the teams who would use the tool. After the internal release, I tracked content through expert revision and asked engineering to integrate the finalised pieces first, rather than wait for the whole batch.
What changed, and what the evidence means.
The baseline came from asking the pedagogy team how long an exercise took. I timed the automated workflow and included a ten-minute allowance for a quick human review. This is an approximate workflow comparison, rather than a controlled productivity study.
The first version was released internally to pedagogy teams in 2025. The broader initiative produced curriculum-aligned questions for English and Chinese programmes. These demonstrate content production and internal delivery; learner outcomes were not evaluated here.
Separate authoring work in 2026
I later extended existing question-generation pipelines to work with figures in uploaded teaching material. Teachers could reuse a source figure or work with an illustrative chart, with separate LLM checks for chart validity and whether a question actually relied on the figure. This was a new capability layered onto colleagues’ existing pipelines.
What I would carry forward.
Prompt quality and review design belong in the same product discussion. For the next iteration, I would track review time, the type of corrections required and repeat usage alongside generation time, so the team could see where the workflow still created work.
Product visuals and interactive examples are recreated for this portfolio with synthetic content. They are not screenshots of learner records or internal systems.
Next case study
Feedback a learner can use