From Playground Interviews to a Working App: Designing a Meal Planner That Agents Built
Challenge
Planning the week's dinners for a family is harder than it sounds. Every family member has different needs, the plan has to turn into a shopping list, and food gets forgotten and goes bad. A study of household meal planning calls it "a complex holistic task requiring constant monitoring of eating, adjustment for tastes and negotiation between people" (Pickering, 2023). For parents of children in daycare there is one more layer: what the children already ate for lunch. Bringing that lunch into the plan was my primary goal. And I picked up this real problem for a second reason: to learn how to bring Claude Code and agents into the design process.
My Role
This was my own project, with no client and no team. I did the user research, the UX and UI design and the product owner's job, and I owned every design decision and the usability test. In the design work I was the art director: intent, taste, direction and the final visual nudges. Claude was the production designer, applying those decisions through the design system, and Claude Code wrote the app. What to build and how it should feel stayed my call.
Activities
- Talked to parents on the playground about how they plan the week's meals, before an agent wrote a line of code
- Desk research on meal planning as a household task, on competing apps and on their success factors
- Synthesised the research into personas, jobs to be done and hypotheses, and picked the core idea: dinners that don't repeat what the children ate at noon
- Wrote a PRD and a design document in one knowledge base, so that research, scope and design rules were shared context for me and the agents
- A product-owner agent, built as a Claude skill, kept the work on track in Linear, with every commit referencing its ticket
- Explored visual directions in Paper, then locked a design system whose tokens and components exist in Paper and in the SwiftUI code under the same names
- Directed Claude Code as it built the native iOS app and about 220 unit tests
- Ran a moderated usability test with families: six sessions
Outcome
A working native iOS app, version 0.1. Its core feature reads the daycare meal plan from a photo or a PDF: on-device text recognition, then Claude parses it into a structured week, and the dinner suggestions avoid repeating the daycare lunch. Around it: a seven-step onboarding that asks only what the plan needs, a weekly plan with a colour per child, recipe import, and a shopping list built from the plan. The app is open source on GitHub. The lesson I took from it: agents are only as good as their context, and keeping that context current is real design work.