The comment was there. The next step was harder to see.
During pedagogy review, feedback could sound mechanical, repeat the same point or open with what a learner had done wrong. There were also inconsistencies between an earlier assessment and the wording produced by a later prompt.
In related progress-report work, a teacher told us the language was too technical for students and parents. I brought that feedback to our designer. Producing more text was not enough; the reader needed to understand what to do with it.
My part in the work.
I reviewed examples with pedagogy, revised lesson-summary prompts and submitted backend changes across three summary formats. I also drafted examples at simple, moderate and advanced language levels, then sought feedback from the domain team.
An engineering colleague integrated the level-based guidance into the wider grading and summary workflows. I worked with our designer on the report experience; the visual design was a collaboration.
What I changed, and why.
Give the learner something to act on.
I changed prompts to lead with a specific strength, use simpler language and frame improvement as a helpful next step. Pedagogy reviewed the new examples and marked what still needed changing.
Keep assessment and explanation consistent.
I flagged cases where a later feedback prompt contradicted an earlier mark. A friendlier rewrite still needed to explain the recorded assessment accurately, so the review covered consistency as well as tone.
Adapt to the reader.
I drafted different language levels instead of using the same register for everyone. The report discussions also led us to simplify the proposed status display and put recommendations into the feedback, following alignment with leadership.
Changes that reached the product.
My revisions to three lesson-summary prompt formats were merged in October 2025. I later added explicit output-language handling. Engineering integrated the level-based guidance into the grading and summary workflows.
Later follow-through, 2026
On another learning workflow, I traced feedback cut off after an abbreviated title to the sentence splitter used by a filter. I fixed that path and worked on summary checks that were rejecting useful advice. My update to the team distinguished the fixed truncation from generic filler that still needed work.
The evidence shows prompt changes and pipeline repairs. It does not establish an increase in learning outcomes or that every feedback issue was resolved.
The quality check needs checking too.
A filter can remove a bad claim and accidentally damage useful feedback. I now look at the whole route to the learner: the source evidence, the generated explanation, the filters and the final screen.
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
Different paths. One course.