Huddle
A two-day experiment in building with AI, and what it taught me about where designers still matter.


Group chat decisions spiral.
Six people, forty-seven messages, nobody remembers what was agreed on. Restaurants, movies, trips, chore splits. The format punishes everyone equally.
Insuring friendships and relationships with our solution.
01
Pick a template
02
Add people
03
Each person fills out their preferences/constraints privately.
04
Huddle proposes one clear recommendation, with reasoning that sounds like a friend who knows the group.
BUILD TIME
2 Hours
BUILt WITH
Lovable, Gemini 2.5 Pro, Supabase
Three design decisions
The reasoning voice is the product.
A cold algorithmic summary is useless. The system prompt became the most important design artifact of the project.
Let the LLM describe entities.
LLMs hallucinate links. The pattern I landed on: the LLM returns structured data (restaurant name, city). The frontend builds deterministic search URLs from it.
Acknowledge what we don't know.
If someone skips a question, Huddle's reasoning says so: "Mika didn't weigh in on budget, so I'm flexing on that."

An introspective
What lovable was great at:
Interactions. I described flows in plain language and the transitions, state changes, and conditional logic just worked. The 4-step create flow took twenty minutes. Wiring Gemini into the results view took one prompt.
Where I pushed back:
Aesthetics. Lovable's default visual language is the same generic SaaS look every AI builder defaults to. I made small pushes on Huddle but deliberately didn't spend the full polish cycle. I wanted to see the default state of the tool honestly.
The conviction I walked away with:
AI tools are fast, but they default to generic. Without a designer with opinions, the output will always read as AI-made. That's the space where designers still matter (for now!)
Lovable gave me a working product in two hours.
Making it feel human would take two days. Here's somethings I'd add to it:
01
More warmth.
02
Illustrations with personality.
03
Microcopy that sounds like a specific writer, not a generic friendly voice.
