I started the way you'd expect: mapping HealthifyMe, MyFitnessPal, and the rest. They were all the same shape: a library of meals, presented, then stepped away from. That gave me a hypothesis, not an answer.
"Design a meal planning app." Broad, and not yet a real problem.
HealthifyMe, MyFitnessPal and the rest all shared one shape: a library of meals, presented, then stepped away from. Passive.
Passive libraries were the problem. A reasonable starting theory, and it turned out, not the real one.
I wasn't testing for this. I was testing navigation with 4–5 people. But the pattern was consistent: people opened the app, looked at Home, and did nothing. Not confused. Not lost. Just not moving.
I pointed testers straight at the meal planner. They still stalled. That ruled out navigation as the cause.
Starting a meal plan requires already knowing what you want to eat, which requires knowing your calorie target, which requires having thought about this before opening the app.
This pattern held across 4–5 informal sessions with people I know. It's consistent enough that I redesigned around it. It is not a validated causal finding: a bigger, more diverse sample might surface a different reason for the same stall.
Once "reduce the cost of the first decision" became the actual problem, there were at least three ways to attack it. I want to be honest about how seriously I weighed the two I didn't pick.
"Users who ate X also liked Y," driven by time of day and static categories.
BenefitTransparent, offline-capable, easy to defend in a design review. Would have shipped fastest.
CostDoesn't adapt to an individual's goal or intake. Narrows options slightly without removing the burden of deciding.
"High-protein week," "Quick meals for busy days." Pick a template, follow it.
BenefitFast to build, predictable, zero black-box risk.
CostDoesn't know what's in your fridge or where your calorie count sits today. Runs out of novelty within weeks. Solves the first session, not the pattern.
Personalized to goal, tracked intake, and available ingredients. Adapts over time.
Why chosenThe only option that makes a decision for the user rather than just narrowing their menu.
CostNeeds onboarding data before it's useful (cold start), and if the first suggestions feel wrong, that trust is expensive to earn back.
I didn't build working prototypes of the other two paths and test them head-to-head. This is reasoned elimination, not an experiment. If I were doing this with real engineering time, the rule-based option is the one I'd actually want to A/B against the AI version, because it's the cheapest to be wrong about.
Not a clean lo-fi → mid-fi → hi-fi progression. The fridge-scan pivot and goal-first onboarding both came out of testing sessions midway through, not the original brief. Each decision, once, with what it actually cost.
The lo-fi Home had a carousel and a content grid with equal visual weight. Nobody testing it had a reason to look at one before the other. This is where the "do nothing" stall was most visible.
An AI Daily Brief card as a full-width hero, everything else pushed below the fold. One focal point, one suggested action.
If the AI's suggestion is wrong or irrelevant that day, there's no visible alternative entry point without scrolling past it. Betting a wrong-but-present suggestion beats no suggestion. Haven't tested the miss case.
The original flow put 10+ inputs in front of the user before any payoff, opening with "Diet Type," a clinical, low-stakes-feeling first question for someone who hasn't decided to commit yet.
A 5-step sequence (Goal → Budget → Dietary needs → Allergies → Calorie target) with a progress bar and a skip option on every screen.
The total information collected is probably similar either way. What changes is perceived effort, not actual effort: a real lever, but not a new insight. It does move value (a first plan) to after step one instead of after ten.
AI suggestions were visible in mid-fi but not interactive: no way to say "not that one," no explanation for why it was suggested. It read as an announcement, not a proposal.
Keep / Swap on every meal, with a one-line reasoning card attached. Full reasoning is available on tap, not by default.
Users who want the full "why" before trusting it get one sentence by default. Betting most people want speed over depth. Weak evidence, since nobody asked to expand the reasoning, but I didn't specifically test for that either.
The earliest version assumed people would manually enter what's in their fridge. Nobody does this. People open the fridge and react to what's there.
A photo scan with a confirmation step: detected items shown as a list, false positives removable, before any recipe generates.
An earlier draft of this case study stated a "30% error rate" as if it were a finding. It isn't. No CV model was built or tested. The confirmation step is the mitigation I'd want in place before finding out what the real error rate is, not proof that it works.
Round one was mid-fi testing. That's where the stalling pattern in Section 1 surfaced, along with the two pivots (fridge scan, goal-first onboarding) that came out of it midway through, not from the original brief. Round two was hi-fi testing, checking whether the redesigned entry points removed that stall. I don't have granular session-by-session notes comparing the two beyond what's already captured in the four decisions above. They tested different fidelities of the same problem, not a before/after of one fixed test, so I'm presenting them as one continuous thread rather than forcing a comparison that implies more rigor than there was. What follows is the tangible record of that thread: the same three screens, at each stage.
With 4–5 testers, those percentages would imply a sample I didn't have. 88% of 4 people isn't a real number.
A clean onboarding funnel, dressed up with percentage points from a sample of four or five people. It would look rigorous. It wouldn't be.
In the original single-form onboarding, most of the people I tested with stalled before reaching a generated plan. In the redesigned version:
[ Kartik: insert your real count here, e.g. "4 of 5 reached a plan without a prompt from me" ]That's a small, informal signal the goal-first pacing removed the specific stall I'd observed. It's not proof the approach scales.
Visible day count on Home. Cheap and common. Present because it's low-effort, not because I have evidence it drives return visits.
A Sunday notification to confirm or adjust next week's plan. A low-effort recurring touchpoint, without demanding daily engagement.
The bet: reducing weekly decision fatigue drives more organic return visits than a notification does. The mechanism I'd measure first.
Every swap is a signal that could improve suggestion quality over time. The mechanism I believe in most, and the one that takes longest to prove.
Early suggestions are the ones users are most primed to doubt. If the first few feel off, the one-liner may not be enough to keep someone from turning the feature off entirely.
MitigationSwap is always one tap away, and every swap feeds the model. That loop needs volume to work, though, and volume is exactly what's missing on day one.
No CV model exists yet to know its real error rate. The confirmation-and-removal step is designed to catch misses before they become bad recipe suggestions.
LimitIt can't fix a scan that misses so often that reviewing it becomes its own chore.
A streak counter and a Sunday notification are both passive. Neither is a strong mechanism on its own, and nothing is designed past week one for when the novelty of "the AI made this for me" wears off.
StatusThis is the least-finished part of the project, not because it looks unfinished, but because it actually is.
The first pass optimized for completeness: every input the app might eventually want, surfaced up front, "Diet Type" leading the way. I went in expecting a navigation issue and found a sequencing one instead, and only because testing forced the reframe, not because I named the constraint up front.
I arrived at the real constraint (reduce the cost of the first decision) through iteration, several days in, rather than naming it before wireframe one. If I'd forced myself to state that north star earlier, I might have caught the Home hierarchy problem and the onboarding sequencing problem in the same pass, instead of finding them separately.
Everything past week one. A streak counter is an easy default, not a real answer to why someone keeps opening the app once "the AI does this for me" stops being novel. That's the harder problem, and I ran out of the 7 days before I got to it properly.
I'd want a real sample before claiming any completion number to a hiring manager. And I'd pick a habit mechanism (probably the fridge-to-recipe loop, since it's tied to utility rather than a reminder) and design two or three weeks past it, not one.
That's not a complaint, it's the point. This case study was built to be skimmed for decisions and read in full for the reasoning behind them. Either way, here's what I'd want you to leave with.
Naming a real constraint from messy signal, then designing every downstream decision around it instead of around a feature list.
Testing early enough that the reframe was still cheap. The Home hierarchy fix and the onboarding fix both trace back to the same four or five people.
Solving for week two and beyond, with a real habit mechanism, and validating the fridge scan and reasoning-card bets with more than four or five people.
Thanks for reading, Kartik