Toddle is a curriculum planning and assessment platform initially built for IB schools, now catering to all major curriculums globally. Progress Reports is one of its most business-critical modules, used by 500+ schools, multiple times a year, to generate printable report cards for students and families across PYP, MYP and DP curriculums.

Teachers have to review assessment data, recall specific moments from the semester, and write a personalised comment for every student in their class. Multiply that by 30 students, across multiple subjects, and you have hours of work happening entirely outside teaching hours. Toddle is built to give teachers that time back. When AI came along, generating comments was the obvious place to start.

In early 2025, we shipped the initial version of Toddle AI for progress reports. Two capabilities:

Voice and output structure: Admins define the tone and format of comments school-wide, so the final report reads coherently whether ten teachers or one wrote it.

One-click auto-generate: The AI pulls in assessment data, teacher notes, class files, and student feedback to produce a comment in seconds.

Teacher response was strong. Adoption was not.

Teachers were willing to try it. They were not willing to submit it. They couldn't verify what the AI had written. If it said a student "showed strong analytical thinking," there was no way to trace that back to a specific piece of work. AI helping them write was a new mental model and they didn't trust it entirely. Two things kept coming up in interviews:

Selective context: Teachers wanted to choose what the AI was working with, not have everything passed in automatically.

References and sources: If the AI made a claim, teachers wanted to know exactly where it came from.

This was not a quality problem. It was a trust problem.

Giving teachers control over context

Earlier the system prompt passed everything in: subject data, learning goal evaluations, assignment ratings. The assumption was that more context meant better comments. What teachers kept saying was that more context meant less control. A teacher preparing a subject comment didn't want assignment-level detail bleeding in.

User interviews revealed a clear split:

• Some schools wanted teachers to have full, manual control over what context went in.
• Others were fine with a default selection, as long as it aligned with what the admin had already configured for the template.

The solution was a layered system: admins configure what data is visible on the report template, and context checkboxes default to match that configuration. Teachers can override. But the default is already right for most of them.

Showing references meant usability tradeoff

Teachers wanted to see exactly what the AI had pulled from before submitting a comment. The obvious pattern was inline citations inside the comment box. The problem: it meant placing an AI interface layer on top of Toddle's native comment editor and it looked like two products colliding.

Technical limitations made inline editing unviable, and speed of execution ruled it out further. The direction shifted to a popup overlay: teachers click to view sources, a focused panel surfaces the references, and they return to the comment editor with full context.

It interrupts the flow slightly. It earns the trust completely.

Reworking the system prompt

Progress reports have multiple comment types: subject comments, comments on learning goals, homeroom advisor comments. Each had its own system prompt, written specifically for what that comment box was communicating.

Once teachers could select their own context, the prompts needed to be restructured. The AI still needed to know what kind of comment it was writing, but now it also had to work with only the context the teacher had chosen, not everything in the system. The prompt logic shifted from "here is all the data, write about it" to "here is what this comment is for, here is what the teacher wants you to connect."

A structural change underneath, invisible to the teacher, but necessary for the output to hold up.

Context selection, smart defaults tied to admin configuration, and the references popup shipped together. Teacher submissions increased. Comments going through the AI layer stopped being rewritten from scratch.

The references UI took the longest to get right. The popup overlay was not the first choice, but it was the right one. Surfacing sources clearly, even at the cost of flow, was what made the feature usable in practice.

Designing for AI adoption is not about making the output better. It is about giving users enough visibility and control that they can own the output. Context is the lever. Trust is the outcome.