Helping families turn a shared cue
into their next question.
A family conversation app that suggests questions from photos or shared topics and connects stories about the people mentioned.
I built an AI-connected prototype, then redesigned how both family members could shape the conversation.
A family wanted to hear more.
Where could the conversation begin?
A Korean-American learner I tutored wanted to hear about their mother’s life before immigration. Their mother was living with dementia, and they didn’t know enough about her life or her home country to know what to ask. The conversation kept circling the little they already knew.
I built the whole loop with Claude.
Upload a photo, generate a question, capture an answer, and build a memory map.
How it works · Illustrative
- Me · 01
Define the interaction
Photo → question → answer → memory map.
- Claude · 02
Turn the flow into a working app
Help implement and debug the React prototype.
- Me · 03
Run it, judge it, request changes
Questions stayed generic. I asked for person-focused follow-ups and more specific keywords.
- Claude · 04
Revise the implementation
Update question instructions and keyword extraction in response to my feedback.
- Me · 05
Retest with the same inputs
I observed more specific keywords. That observation did not establish better follow-up questions.
The AI recognized the photo,
but its questions stayed generic.
The answer mentioned someone bringing flowers. The AI stayed with the object instead of exploring the person and the situation.
“Do you like flowers?”
The AI asked simple questions.
“Who brought the flowers?”
“Why did they bring them?”
Ask who was involved and why, using the answer’s context.
Reconstructed examples show the requested direction. I retested the same photo and answer and observed more specific extracted keywords. That did not establish better follow-up questions.
Feedback pushed me to rethink each person’s role.
I demonstrated the prototype to five designers at junior, mid-level, and senior levels. Their feedback prompted me to reconsider each person’s role.
“What does the older adult do while their guardian is typing?”
Let the storyteller choose the starting topic
I added physical topic tiles so the storyteller could choose what to talk about. A photo became optional, opening a way to begin with a person, place, or time.
- Family memberUploads a photo
- Family memberAsks a suggested question
- StorytellerAnswers
- StorytellerSelects topic tiles
- Family memberScans and asks
- StorytellerShares
After the first round of feedback, I designed V2 around the storyteller choosing physical topic tiles. The family member would scan the tiles and guide the conversation. Tile recognition was simulated.
“Is the granddaughter only there to type?”
Let both people choose the starting topic
After seeing V2, a senior designer and a colleague asked whether the granddaughter was only there to type. I added shared topic selection so she could suggest what to explore, while still managing questions, recording, and saving.
Start together.
Build a conversation to return to.
Start with a photo
Bring a family photo into the conversation. Look at it together and choose what to ask.
Start with tiles
Choose topics together when starting without a photo. Either person can suggest a person, place, or time.
Chooses what to share
Can change or decline a topic.
Asks and operates
Chooses topics with the storyteller, asks questions, and manages recording and saving.
AI support Photos or topic cues + answers → suggested questions.
Saved answers → memory details and summaries.
Follow the answer into the next question
The family member asks an AI-suggested question or one of their own. The storyteller answers aloud; voice input or typing captures the response.
Find related conversations through the Memory Map
AI extracts keywords from the conversation into the Memory Map. Select a person, place, or time to revisit related conversations and an AI summary.
The current prototype stores conversations in this browser. Cross-device family syncing and physical-board sensing are not connected.
What building an AI product taught me
Recognizing the photo was not enough
The AI recognized flowers, but “Do you like flowers?” moved away from the story. I learned to use the answer—not just the image—to guide questions about who brought them and why.
Shared topics did not remove the operator role
V3 lets both people choose topics, but the family member still manages questions and recording. I learned to distinguish sharing control of the conversation from sharing control of the app.
Building with Claude let me move from an idea to behavior I could inspect. I learned to give specific feedback, rerun the same inputs, and distinguish what I observed from what I hoped would improve.