Role: Solo Product Designer + Builder
Timeline: 5 hours (May 2026)
Status: Live & iterated
Stack: Claude AI, MCP, Next.js, Supabase
After relocating to the U.S., I fell in love with cooking. But remaking dishes meant searching through old AI conversations. So I built an AI Chef Agent—a system where Claude handles all the complexity (discussing, generating, managing recipes), while the app stays simple and read-only.
This is the beginning.
<aside> <img src="/icons/robot_yellow.svg" alt="/icons/robot_yellow.svg" width="40px" />
My request: I want to make Taiwanese Beef Noodles!
AI workflow: I discuss the recipe in Claude → Say "Save" → Recipe auto-saves through MCP → Open Cici Table → It's there.
</aside>
I used video to demonstrate my working prototype. I built the agent in my personal Claude account to quickly test the AI-human workflow.
<aside> <img src="/icons/apron_yellow.svg" alt="/icons/apron_yellow.svg" width="40px" />
The read-only recipe app: https://recipe-app-o78n.vercel.app/ (passkey: cicirecipe)
Note: This is just the display layer. The AI Chef Agent (the actual intelligence) lives in Claude with system prompt + MCP.
</aside>
I'm a beginner home chef, and I save recipes from social media. But the cooking experience is fragmented—I can't find recipes during meal prep because the platform wasn't designed for read-and-cook workflows.
To improve this, I use AI (Claude, GPT, Gemini) to discuss, generate, and customize recipes. But every time I want to remake a dish, I have to dig through old AI conversations.
The gap: I didn't need another recipe website. I needed a personal kitchen assistant that remembers my preferences and makes past recipes instantly accessible. And more, in the future, to build a seamless cooking experience.
I started building Recipe Copilot with embedded AI: chat interface, manual editing, step-by-step cooking mode.
Then I realized what was happening: I was spending time on recipe editing, management, demo AI flows—features I couldn't validate without real usage. Every change to the workflow required touching the whole stack. Too much infrastructure meant slower iteration.