Team Null 三人如何在六小時內不寫一行程式碼，贏得 Grok Bot Tokyo Hackathon

We humans didn't write a single character of code. We barely touched our keyboards. We just talked to AI agents, basically thinking out loud.

And we took first place at the Grok Bot Tokyo Hackathon.

Here's how that day went.

The half-asleep post on X that turned into this write-up

This is the English version of my original Japanese post on note.

Morning: Arrive late, find a team

I woke up and headed to the venue alone, without really thinking about a team. Then I stopped for meat udon at Oniyanma in Ochanomizu and ended up late.

When I got there, N (@_I__l__l__l__I_) was sitting at a table by himself, looking at me like he wanted to join my party. So I sat down next to him.

The morning was a beginner-friendly hands-on session. Connect to the venue Wi-Fi, get $50 in credits... except I already had a Cursor subscription, so I couldn't redeem them. Oh well. I set up a fresh environment instead.

About 30 minutes later, Limo (@vibe_walking) showed up even later than I did. When he signed up, he'd been assigned to a random team with people he didn't know.

He quit that team in one minute. lol

I asked N next to me to team up, and with Limo fresh off his one-minute stint, we became a team of three. Limo suggested the name: Team Null.

The newly formed Team Null. on X

Somehow I ended up as the leader and took on direction. Honestly, that's easier for me. I'm the type who loses power when someone else tells me what to do.

12:45: Fifteen minutes to pick an idea

Limo and N went out to the convenience store. I'd already stocked up at the FamilyMart on B1, so I didn't go with the team. Instead, I remembered Cue had a phone number, and had Cue's AI agent call Limo to ask him to bring back a Diet Dr Pepper.

Limo's reaction: he brought me a Red Bull and got a call from an AI agent. lol

The three of us ate the boxed lunches at the venue and did some very informal team building. I'd met both of them a few times, but we'd never built anything together, so I didn't really know who was good at what. We started by chatting.

After lunch, we made an X DM group with the three of us and Naru-Ebi-chan, my AI agent. From then on, that group was our notepad. Limo recorded our call in the DM group, so the entire planning session was captured as audio.

We gave ourselves about 15 minutes, until 13:00, to come up with the idea. I set an alarm on my iPhone. No endless brainstorming.

There were three judging criteria, so we used those to guide the discussion. During lunch, Limo had mentioned recently trying to order through an agent and real-time voice. That sparked ideas like "AI places the order" and "when your beer mug is empty, AI orders another one automatically."

After 15 minutes, we had the broad plan:

• Build a mock restaurant on an iPad, with something that looks like a restaurant ordering tablet.
• Build a mock backend to receive orders.
• Let Grok Bots talk things over with one another.
• Address real-world issues like food dislikes and allergies.

At that point, I could already see where we were headed.

13:00: Split the work, and start with the slides

We gave the planning-session audio to Grok Bot and had it transcribe and organize everything. Here's how we split the work:

• Me: Slides, direction, and timekeeping.
• Limo: Frontend and UI.
• N: Backend, the restaurant-side server.

Limo set up a private GitHub repository for the team.

I went straight into making the slides. Starting with the presentation helps us:

• Get everyone on the same page about what we're building and where we're going.
• Make the target audience and concept clear.

To me, it's more about drawing a map for ourselves than making something for the judges.

I put just a rough outline into Google Slides and connected Grok Bot to it. After that, I could throw rough instructions at Grok Bot and it would turn them into something presentable. It already had the planning-session audio, so it knew the context.

Giving the hackathon-master Grok Bot editing access to Google Slides

Limo was firing off voice instructions to Grok Bot at ridiculous speed. N was doing the same through the DJI Mic he'd brought along. Nobody was typing anymore.

Just three people endlessly talking to themselves. A very weird little corner of the room.

N's post about all that voice input

14:00: Integration, and the moment the bots connected

With clear roles, everyone was moving ridiculously fast. We'd set up the AI agents to read our X DM group, so in a way, that group became the team's communication protocol.

We let everyone work for an hour without doing any integration testing. When the 14:00 alarm went off, we checked in.

By then, we had the slides, a backend stub, and the frontend, minus the menu photos. After a bit of back and forth, the frontend and backend were connected on Cloudflare.

I used Grok Bot to create an organizer bot and three persona bots, one for each person.

My post about creating the bots

Then I set the next alarm for 15:00, and we went back to work. Scanning a QR code now brought up the restaurant menu on a phone. I tried various ways of getting the Grok Bots to communicate and eventually settled on SKILL + Webhook.

The development repository Limo had set up looked like it might be getting chaotic, so I created a separate repository for the submission.

15:00. Thirty minutes until our team's internal deadline.

Some parts still weren't working, but we'd implemented the concept and the bot-to-bot communication. When our Grok Bots actually connected to each other, everyone went nuts. We could feel it: this might work.

What remained was fixing bugs and making the UI easier to read. Limo took over some of the backend work that was getting hectic and got it finished. Pull requests flying everywhere. lol

Continue to Part 2: AI judging, the pitch, and first place

What we built (GitHub)

The presentation from the day