“I have no idea how this app can help me .”
Habit building app
Users didn't understand how the app would help them eat healthier?
LUMOSTEP
AI-POWERED FOOD GUIDANCE
Making healthy eating easier, one meal at a time.
An AI food companion that helps busy vegetarian professionals make small, achievable improvements to the meals they already eat.
Three versions, one question: what actually helps people eat better?
V1
Habit building app
Found the wrong problem
V2
WhatsApp AI MVP
Found the right behavior
V3
AI-native mobile concept
Translated MVP learnings into an interface.
V1-HABIT BUILDING APP
My first solution was solving the wrong part of the problem.
I designed LumoStep around food education, small goals and weekly plans. It encouraged healthier habits without accounting for the circumstances that shape people's meals.
User testing surfaced three problems.
“How can I get my questions answered?”
Missing conversation
They wanted to ask questions about meals and recipes.
“My family decides the food that we eat, not me”
Existing family routines
Changing food choices often meant convincing the entire household.
CHALLENGE
How might I help people make healthier choices within the meals and routines they already have?
V2-WHATSAPP AI MVP
I tested the AI before designing the interface.
To test whether personalized guidance could fit into existing food routines, I built a working WhatsApp AI MVP in one week, focusing on Lumo's behavior before designing a new interface.
Conversation solved one problem and exposed three more.
I tested the working WhatsApp AI MVP with five people. Two in observed sessions and three independently. The testing surfaced three usability problems.
“Do I need to write everything?”
Too much input effort
Typing out meals and ingredients was tiring.
“Too much text, I don’t feel like reading.”
Too much output
Responses were too long
to act on.“I don’t know if I can follow this diet?”
Too much change
Recommendations asked for more than people could realistically do.
THE INSIGHT
Useful guidance has to be both achievable and easy to access.
Turning insights into AI behavior
I rewrote Lumo's behavioral guidelines in Cursor (lumo-guidelines-v2) around three rules.
01
One change at a time
One achievable improvement to an existing meal is the default.
02
Reduce irrelevant steps
Reuse stored context instead of asking again.
03
Faster clarification
Photo input and quicker clarification, carried into the next interface.
A key decision: when the AI's suggestion wasn't the right one
During AI exploration, an alternative was to recommend immediately from available ingredients.
AI PROPOSED
Recommend a recipe immediately from available ingredients, to cut down on questions
I DECIDED
Pogressive personalization. Use existing context, and ask only when a question materially improves the recommendation.
Why: speed matters, but a fast recommendation isn't useful if it ignores the person's context.
LUMO V3AI-NATIVE MOBILE
Designed V3 collaborating with Claude
I used Claude to generate and critique interaction models, then designed the final experience in Figma. The central question: when should Lumo ask, and when should it recommend?
These three decisions shaped the experience:
Combined the context strip with image input to reduce effort and make information easier to scan.
Moved context chips closer to the recommendation.
Placed context after the recommendation to make the AI's reasoning visible.
Explored three entry patterns to make Lumo's capabilities clearer. I chose two guided starting options so users could begin with either their meals or available ingredients.
V3KEY DESIGN DECISIONS
01
From typing ingredients to showing what's available
V2 testing revealed that describing ingredients through text took too much effort. I introduced photo input with quick clarification when Lumo couldn't confidently identify an ingredient.
- Share ingredients with a photo
- Clarify only when needed
02
From lengthy AI responses to scannable recommendations
V2 testing showed that long responses were difficult to follow. I designed a recommendation card that shows one small change first, with recipe details available when needed.
- Scan the recommendation
- Expand only when needed
The complete experience
REFLECTION
AI can generate possibilities. Designing the right behavior takes judgment.
Working with AI taught me that generating options is easy; evaluating them against real user needs is where design judgment matters most.
01
Design judgment over AI output
AI helped explore possibilities, but I evaluated each direction against user needs and product constraints.
02
Test AI behavior early
Testing real conversations revealed problems that polished interface explorations couldn’t expose.
03
Separate design from validation
V3 translates what I learned into a new interface, but user testing is still needed to validate those decisions.
WHAT’S NEXT
Test whether photo input reduces the effort of sharing ingredients, and whether focused clarification helps users reach useful, achievable recommendations with fewer steps.