Mar 17 2026☘️
Before this class, I thought behavior change design was about identifying a problematic habit and then building a tool that made the healthier alternative more accessible. I assumed the hard part was engineering and interface polish, and that if you put the right option in front of someone at the right time, the rest would follow. I seriously underestimated how social context, identity, and emotional self-regulation are woven into even something as ostensibly straightforward as what a college student eats at midnight. Working on LunaCart challenged these assumptions throughout the 10 weeks of 247B. We were initially surprised by the baseline study; we went in expecting students to say they wanted to stop eating late at night, and instead they told us the behavior itself was fine, even valuable. Instead, the food available to them was the problem. That pivot forced us to reframe our research midway through synthesis, which proved disorienting but ultimately produced a much more adept intervention. I loved the grounded theory process for this reason: coding raw interview transcripts and watching subtheories emerge from participants’ own language gave the project a foundation that felt earned instead of feeling assumed. The journey maps and proto-personas were also generative; translating data toward narratives made our users feel accessible in a way that demographic profiles alone never could.
What I found less useful, at least for our particular problem space, was the assumption-testing phase as we implemented it. The Strategyzer test cards imposed a clean hypothesis-test-measure-criteria structure that worked well on paper, but our tests were so time- and sample-size-constrained that the results felt suggestive rather than conclusive. The leaderboard test, for instance, confirmed that social comparison motivated task completion, but the artificial context made it hard to know whether that would transfer to real late-night food decisions. If we were to redo this, I think we would invest more time in longer, more naturalistic assumption tests, even if it meant testing fewer assumptions overall.
A specific problem we encountered was the menu bottleneck during the intervention study. We partnered with One Step Foods, expecting that free healthy snacks would drive adherence, but participants told us the options did not match their cravings, and one participant said she would rather pay for something she actually wanted than receive a free item she did not. This was humbling. It revealed that our team had substituted our own logic (free = desirable) for the user’s actual value system (taste and fit outweigh cost). We resolved this by reframing the menu in the final prototype as a curated, flexible browsing experience rather than a fixed set of prescriptive options. I will personally say that the tension between curation and personalization remains unresolved for me (more to think about!).
On ethics: the pre-commitment mechanism is the feature I think about in greatest depth. It works by using a temporal gap between intention and temptation, which is a well-established nudge, but the line between supportive scaffolding and paternalistic control is thinner than I appreciated. For most users, an optional morning pledge that they can modify or ignore is autonomy-respecting. For a user with disordered eating patterns, however, the same mechanism could reinforce rigid food rules or trigger guilt when they deviate from their plan. We tried to lessen this by keeping the commitment checkbox optional and framing the morning log as reflective rather than evaluative, but a production version of this app would need more robust screening and opt-out pathways. Privacy is another consideration: logging what someone eats, when, where, and with whom generates a detailed behavioral profile, and the social accountability features would make portions of that profile visible to peers. We would need clear, granular consent controls and a policy that ensures eat-rate and streak data are never surfaced to institutional actors, such as dining services or student health entities.
This experience broadly changed how I think about design work. I now see that the most important design decisions happen before a single screen is sketched: in how you frame the target behavior, whose inputs/data you center while synthesizing, and which assumptions you prioritize for testing. These upstream choices constrain everything downstream, and they deserve as much rigor as the pixel-level decisions that feel more tangibly like “design.” Next time I face a similar project, I will push harder to let user data reshape the problem definition early, resist the temptation to lock in an intervention concept before the baseline findings are fully synthesized, and build longer feedback loops into assumption testing so that the evidence guiding the final prototype is as robust as possible.
