AT
Alexandra Tragut
AI Lesson Generator
Helping teachers turn ideas into editable lesson drafts
I designed an AI lesson-planning experience that connects generated lesson drafts with evulpo’s existing learning content. Research with teachers shifted our focus from comprehensive planning toward flexible drafts that teachers could review, adapt, and make their own.
My Role: Primary Product Designer, responsible for research, interaction design, UI design, and UX writing
Team: Three developers
Timeline: Three weeks of design, followed by six weeks of development
Status: Launched September 2025
The Challenge
Teachers need lesson plans that fit their class, teaching approach, and available time. Generating text is only part of that work: they still need to evaluate the output, find suitable learning materials, and adapt the lesson.
Our goal was to help teachers move from an initial idea to a useful draft within evulpo. The experience needed to support different levels of AI familiarity, follow a relevant pedagogical structure, and connect teachers with existing content.
With three weeks for design, I focused research and prototyping on two questions: how should teachers describe what they need, and how should they work with the result?
Research: teachers wanted a starting point
I interviewed four lower-secondary teachers (Sek I, grades 7–9) to understand their planning habits and experience with AI.
The teachers we spoke with described using AI for inspiration and help getting started. They wanted ideas they could shape around their own classes, rather than a tool that made every planning decision for them.
This finding gave us a design principle: make it easy to create a draft, while keeping teachers in control of what happens next.
Decision 1: offer guidance without restricting experienced users
Selection
Dropdown
Chat Style
I compared selection-based, dropdown, and conversational input concepts using low-fidelity prototypes. Nine out of ten participants in a teacher poll preferred the conversational approach, describing it as more flexible and collaborative.
I then developed two chat variations:
-
Guided mode: suggestions and step-by-step questions to help teachers describe their needs.
-
Direct prompting: open input for teachers comfortable writing their own requests.
In comparative sessions with four teachers, participants preferred the speed of direct prompting. However, this group consisted of early AI adopters, which limited what we could conclude about less experienced users.
We chose to support both approaches, with an onboarding question directing teachers toward a starting mode. Guided mode addressed a need we anticipated in the wider audience. With more time, we would have tested guided mode further with teachers less familiar with AI.
Alongside the interface, I defined the information needed to generate a lesson: grade, topic, duration, and lesson type. These inputs also supported the logic connecting generated plans with evulpo content.
Guided Mode with suggestions
Direct Prompting Mode
Decision 2: focus on individual lessons
We explored generating multi-day schedules, but feedback challenged the value of that feature for the teachers we tested. They described adapting plans as they taught and using AI for specific ideas rather than following a fixed sequence.
We stopped development of multi-day planning and focused on a single lesson draft. This narrowed the scope around a more immediate need: helping teachers find useful material and adapt individual parts of a lesson.
Earlier design with multi-day option
Final design
Decision 3: make the output easier to review and adapt
The initial output placed the conversation beside the generated lesson, allowing teachers to refer back to their requests. Testing also highlighted a problem with the lesson itself: long, linear text was difficult to scan during preparation.
Break the lesson into manageable sections
I reorganized the output into modular, collapsible sections and added persistent navigation. Teachers could move directly to a relevant part of the lesson instead of reading the entire response from top to bottom.
Before
After
Use a familiar teaching structure
I structured the lesson around the PADUA pedagogical model discussed during the project: Problem, Build-up, Work-through, Practice, and Apply.
I aligned the interface and system prompts around these stages to make the generated draft easier to relate to teachers’ planning practices.
Integrating teaching structure
Frame the result as a draft
Some participants hesitated at the idea of “saving” an unfinished plan. They saw the output as material to explore, rather than a completed document.
I introduced language such as “draft” and “inspiration” to communicate that the result remained open to revision. The intention was to encourage experimentation without suggesting that generated content was ready to use unchanged.
Modifying and saving from the header
Editing individual sections
The final experience
The design brings conversational input, structured lesson drafts, and evulpo learning content into one flow.
Teachers describe their lesson needs through guided or direct input, generate a draft, and review its sections and suggested materials. The draft provides a starting structure while leaving room for teachers to adapt the lesson to their class.
Responsive Designs
Outcomes and next steps
The AI lesson planner launched in September 2025, with a scope focused on individual lesson drafts.
The direction was shaped by teacher research alongside substantial input from senior leadership and sales, who brought their perspectives on teacher needs. Within the tight timeline, we could not conduct further research to validate all of those assumptions.
With more time and development capacity, I would have tested the resulting experience with a broader range of teachers, particularly those less familiar with AI, and iterated on the guidance and lesson-editing flow. Other product priorities took precedence, so that follow-up work did not happen.
What I learned
The most consequential design decision was reducing the scope. Multi-day planning offered more output, but the teachers we tested placed greater value on ideas they could adapt immediately.
The project also highlighted the importance of distinguishing observed user needs from stakeholder assumptions. Leadership and sales contributed perspectives that informed the direction, but we did not have enough evidence to treat every proposed need as validated.
In future projects, I would make that distinction explicit early: document what we know, identify the assumptions most likely to affect the experience, and agree on which need testing before development. When timelines limit research, that clarity helps the team understand the uncertainty behind its decisions.
I also learned to plan validation as part of delivery. We intended to keep iterating, but shifting priorities left important questions unanswered. Agreeing on a small, focused validation step and the capacity to act on it would give that work a clearer place in the project.