Inspiration
NextUp was inspired by something we kept hearing from neurodivergent students, parents, and educators: the hardest part of the day often isn't the activity itself; it's switching from one activity to the next. Many existing tools focus on schedules or reminders, but they don't guide students through the actual transition in a predictable way. We wanted to build something that breaks transitions into small, manageable steps while adapting to each student's needs, making change feel less overwhelming and more predictable.
What it does
NextUp generates short, predictable, step-by-step scripts to help neurodivergent K-12 students move from one activity to the next. This transition is often harder than the activities themselves for some people. A parent or educator will build them a profile that saves their name, type, preferred tone, and any notes as to what helps them with transitions. NextUp returns a script that’s broken into individual steps, each shown one at a time with an icon, a live countdown timer, and a read-aloud option. If a student needs more time before transitioning, they can press a button to adjust the timer so they aren’t forced to finish early. Every script is realistic: it never invents other people’s actions or promises (like a teacher giving a reward) unless the student’s profile says so, and it explains why when something’s different from the usual routine instead of just announcing the change.
How we built it
NextUp’s backend is a lightweight Java-based server (no external frameworks, just the JDK’s built in HttpServer and HttpClient) that calls the Anthropic API with a very detailed prompt, asking for a script that’s broken into a greeting, an optional reason for any change in the routine, individual steps that each have an icon, and a closing line. The frontend is plain HTML/CSS/JS across four pages: Home, Profiles, Studio, and Library, which profiles and saved scripts stored in the browser via localStorage, so no database was needed. The generator renders the AI’s response as a step-by-step walkthrough with a countdown timer, progress dots, and a read aloud option (the Web Speech API).
Challenges we ran into
One challenge we ran into was getting our AI-generated script to perform the way we expected because it kept making small but well-meaning assumptions that weren’t necessarily true. In the early versions of our program, the AI would write as if it would personally hand the student their next material (e.g. “I will hand you your math worksheet”). In the later versions, the AI would assume there was always a teacher or parent in the room with the student and would give them tasks, like rewarding a student with candy. The student’s profile never actually said either of those things. The AI simply guessed what was missing because it seemed plausible. It’s not usually a big deal for a general purpose assistant, but it is for a tool like NextUp that’s supposed to be predictable, especially for someone who depends on that consistency. When the script makes promises it can’t keep, it can undermine the trust a user has in NextUp. To fix this, we had to be explicit in our prompt that the AI should never assume who is present in the room or what they’ll do for the student unless the student’s profile says so. It also highlighted an important lesson: for this audience, it’s better to be precise and accurate than to make the language sound reassuring.
Accomplishments that we're proud of
We're especially proud that our tester's opinion changed as the project improved. Early on, they pointed out that the scripts sometimes made unrealistic assumptions about what teachers or parents would do, which made the experience feel less trustworthy. After redesigning our prompts to remove those assumptions and make every script more predictable, they told us they really liked how clear the step-by-step format was and especially appreciated being able to adjust the timer instead of feeling rushed. Seeing feedback directly influence the product and then hearing that those changes made the experience better was one of the most rewarding parts of the project.
What we learned
We learned that designing for neurodivergent users requires a very different mindset than building a general AI application. We initially focused on making the AI sound natural and encouraging, but through testing we realized that consistency and predictability mattered much more than creativity. We also learned how much prompt engineering affects reliability; small wording changes could completely change whether the AI made assumptions or stayed grounded in the information it was given. Beyond the technical side, we gained a deeper appreciation for the value of iterative user testing and designing with users rather than just for them.
What's next for NextUp
Our main goal is moving profile and script storage from browser-based local storage to a shared backend, so a student’s profile will be saved across multiple devices and can be accessed by both parents and educators rather than being tied to a single browser. We also plan to let students and testers mark which specific step wording, pacing, or icon choices worked best for them over repeated use so the scripts adapt to what’s actually effective for that individual. In the longer term, we’re interested in expanding beyond text and icon based steps into visual schedule boards with calendars, giving teachers a way to prebuild a full day’s transitions in advance rather than having to generate them one at a time. Finally, we want more neurodivergent students to test our program. We want to go from one tester during the hackathon week to an ongoing feedback process that will shape the future iterations of the tool.
Log in or sign up for Devpost to join the conversation.