This 7-month capstone project was a collaboration with the Open Learning Initiative (OLI), an open-source platform that applies learning science and technology to improve student outcomes. Within OLI's REAL CHEM course, our goal was to make AI support more visible, timely, and useful — helping students get help at the moment it matters, without interrupting their learning flow.
Although DOT (Digital Online Tutor) was already embedded throughout the course, students often overlooked it or encountered it at the wrong moment. As the team's Product Designer and Prompt Engineer, I designed DOT AI Activation Points — a contextual prompting system that surfaces DOT at key learning moments. I defined the activation strategy across three levels, shaped prompt behavior using learning science principles, designed student-facing interaction patterns, and built reusable UI components for the DOT conversation experience. I also designed onboarding materials to help students understand how and when to use AI support. Validated with 373 REAL CHEM 1 students in Summer 2025, the system raised DOT interaction from 44% to 80.4% and scored ~85 on the System Usability Scale.
Contextual AI prompts drove 941 interactions across 373 students in Summer 2025 REAL CHEM.
Increased in DOT awareness through onboarding and contextual prompt strategies.
DOT interaction rate increased from 44% to 80.4% after activation.
373 SUmmer 2025 Real Chem Students
We observed that students often needed help but weren't seeking it at the right moment, which led to low engagement with support overall. That shaped our initial hypothesis: If support became easier to access in the learning moment, engagement might improve.
We considered two directions, then moved forward with DOT as an in-course AI tutor — it could provide contextual support directly inside the learning flow. Since it already knew what page the student was on, students could ask page-specific questions without reconstructing context somewhere else.
Q&A · quizzes · instruction
Contextual support within the learning flow
However, When we tested this with 15 students, we found that many still overlooked it because it remained a passive icon in the bottom-right corner, and students still had to initiate the interaction themselves.
How might we design AI support that reaches students at the right moment, without interrupting the learning flow?
DOT activates on a student's first visit to a page, orienting them before they begin.
🟢
DOT appears below a specific paragraph when the student clicks its AI icon — targeted support on a single concept.
🔴
DOT activates at key activity moments, such as an incorrect answer or a hint request — where students are most likely stuck.
Page Level Activation
Activity Level Activation
Paragraph Level Activation
Authoring Page Level Trigger
Authoring Activity Level Trigger
Authoring Paragraph Level Trigger
If prompts are shaping the student experience, what makes a prompt effective in this system?
We grounded prompt design in real learning difficulties rather than generic AI writing. We worked with a chemistry expert to identify where students were likely to get stuck, then translated those sticking points into targeted activation logic and prompt behavior.
That meant prompts weren’t one-size-fits-all. Each one was tied to a specific learning challenge and shaped by learning science principles, like guiding reasoning, surfacing misconceptions, or prompting reflection.
Identified real student sticking points with a chemistry expert
Translated each learning difficulty into a targeted activation and prompt
Shaped prompt behavior with learning science principles
If we had more time, there are two things I’d want to test next. First, I’d want to understand which activation moments drive the most meaningful engagement, not just the most clicks. Second, I’d want to compare prompts more deeply in terms of trust and reasoning quality.
The biggest takeaway for me is that this project taught me to think about AI as behavior design, not just answer generation. I turned AI from a passive feature into a workflow, and built a more scalable intervention model through timing, prompts, and UI.
.png)
