Case study / Applied AI, own brand
Built with a wary frontline user, one session at a time
My wife is an activity coordinator on a memory care floor. Most of her day goes to leading groups, and most of what is left goes to prep: games to print, exercise classes to plan, supplies to track. Bloomday takes that prep on. The interesting part was building it with someone who was wary of AI.
Live; first use at work is next- Discovery with a wary user
- Judgment on what to remove
- Production Claude engineering
The engagement
Challenge
A real need, and a user who was wary of AI
Activity staff in memory care run group after group with people whose abilities change from day to day, and the prep never stops. My wife is one of them. Like most people asked to try a new system, she was wary at first, especially of AI, and our first session turned into a list of what wasn't working for her. That list was exactly what I needed.
Action
Cut the demo short, build her list, show it back
I stopped the demo and took notes instead. Her list included things I had been proud of: tips on how to talk to residents felt condescending, the craft ideas were mostly her setup, and a timed exercise class could not follow her music. I built her fixes quickly and opened the next session by showing her own ideas working. That got her curious. She started suggesting more, and at the session after that I could show her fixes and her new ideas together.
Result
Enthusiasm first, adoption next, measured when it happens
The person who was wary of it now sends ideas: trivia, conversation cards, coloring pages, word searches. Most of them are built, at four difficulty levels so the material fits the group in the room. She has not used it at work yet, and this page does not claim she has. The app counts what gets used and estimates hours returned, labeled as hypotheses until real use measures them.
How it is built
Listen, build, check, hand back
Every AI response is checked on the server before she sees it, and a person approves anything that goes to print.
- 01Judgment on what to remove. I removed a resident-tracking feature entirely, because a personal project should not hold anything about the people in care, even anonymized. I also took out the staff coaching she found condescending. Bloomday is an activity builder and nothing more.
- 02Her ideas, shipped. Printed move cards with stick figures she can glance at while leading, a one-page class sheet, seated tai chi and yoga that are actually different from each other, picture bingo with a phone caller mode, trivia without scoring, conversation cards, sudoku, crosswords, and coloring pages.
- 03Code where code is better. Puzzle grids come from algorithms; Claude only supplies themed words and clues. Coloring pages use OpenAI's GPT Image 2, picked in a five-model bake-off, because Claude does not draw.
- 04Production habits on a side project. Developed with Claude Code as a Next.js app on Vercel. Strict tool schemas with server-side validation, 1,200+ automated tests, CI, and a live check against the real model before every model change. That check caught four regressions on the move to Claude Sonnet 5 and three more on Sonnet 5.5, all missed by the mocked tests and fixed before release.
- 05Documented like production work. A runbook, a Responsible AI data inventory, and a testing guide. The activity library is also packaged separately as an MCP server and a Claude Agent Skill.
What it proves
Her doubts became the best source of requirements. What won her over was seeing her own ideas built and back in her hands at the next session.
See the rest of the work →Built around a family member's work on a memory care floor. Nothing about any person in care is stored or shown, and the facility is not named. Time savings are not claimed because they have not been measured yet.