A multiplayer, no-login web app that settles where a group should eat. One person shares a code, everyone votes from their own phone, and the app reveals a winner with real restaurants nearby.
If you pick and the food is bad, it’s your fault, so the default is “anything is fine” and nothing gets decided. EatWhere? hands the call to the group vote, then to the algorithm. It’s nobody’s fault.
Anything that needs an account, a download or more than 30 seconds to explain loses to the WhatsApp thread. The bar isn’t “better than other apps”. It’s “better than the group chat”.
Most group-decision tools surface a list and leave the hard part to the group. EatWhere? ends with a winner, announced to everyone at once, with real restaurants ready to tap.
Most group-decision tools end with a list. EatWhere? ends with a moment, and everything before it is build-up. The reveal has three states.
“I spent more time designing that 3-second screen than anything else in the product.”
No account and no app store: a name, a four-character code and a shared link. If the product adds a step, it loses to the group chat. The landing page does one job: Start a session, with join-with-code beneath it, and nothing asks you to sign up. Sessions are single-use and stateless; nothing persists after the reveal.
Nine in ten people in a session only see four screens: join, vote, confirm and reveal. The host dashboard matters, but later. V1 was built entirely around getting the participant loop right.
Voting is the core interaction. Pick up to three cuisines from a 3×3 grid, each card a photo, emoji and name; once three are chosen, the rest dim. The button names the choice: “Vote for Japanese”, or “Vote for Italian · Thai”.
Vote Up to three cuisines, price range and dietary needs.
Reveal The winner, announced to the whole group at once, with up to three nearby restaurants ready to tap.
Restaurant Opens straight in Google Maps.
Four characters, shown large, with the full invite link and a copy button. No O, I, 0 or 1, so nobody misreads it.
Joined, voted and a countdown, updating in real time without a refresh.
Everyone’s name with a waiting or voted badge.
Same voting UI as everyone else, which disappears once the host has voted.
The AI proposed almost the whole visual direction, with light direction from me. The structure validated quickly, but it felt dated in front of real users. I’d let the tool lead too much.
I started from references and a defined energy, and asked for distinct directions, not variations. A dark, moody option was too busy for a 30-second experience; the bold, high-contrast one won. I shipped it knowing it wouldn’t please everyone: bold condensed capitals in red and lime. Warm, not fast food. Playful, not childish.
The whole product was designed and coded with Claude. No engineers, no agency: brief, flows, UI and front-end code, all directed through AI with me making the calls. It isn’t a story of AI doing the work. It’s about directing it, building judgment along the way, and being honest about where it broke down.
Tell me about the role or the project.
