Ambition for Korea, Now Open Source

I'm putting the way we ran the Ambition for Korea hackathon on GitHub as open source.

bloom-hackathon-kit is a kit for running a hackathon where teams build in one day and get judged by VCs that same day. It's the actual code we used at Ambition for Korea (Bloom × Anthropic, October 24, 2026, at the Seoul AI Hub).

At the event, I kept hearing that the way we ran it was unique and like nothing people had seen before. The funny part is that I've never actually taken part in a hackathon. We've done plenty of hands-on sessions at Bloom events, but those were closer to buildathons. This time, each of the eight finalist teams stood to win $5,000 in Claude credits, and VCs had carved time out of busy schedules to be there, so I couldn't be careless about it. If the operation had been sloppy or something had gone wrong, we'd have had a fairness problem, I'd have let down our partners, and the energy in the room would have died.

So the day before, I dumped all my worries and plans on Claude and we built it together. The ops console, the live dashboard on the LED screens at the venue, the VC judging screen, the participant submission form: I made all of it with AI and used it on the day. I didn't know we'd need this many screens either, but I decided to trust the AI and hand it over.

The rules

Above all, this was a survival format. The key was to keep the rules simple and still bring out people's competitive streak.

1/ Building starts at 10 a.m. If people build ahead of time, fairness goes out the window. So we revealed nine missions written by the VCs that morning, and each team picked one to build. We kept the topics broad but made teams work on problems the VCs actually cared about. Everyone also got the same resources: $500 in Claude credits per person.

2/ All development happens in a brand-new public repo. Your code and your commit timeline are out in the open. Some people worried their ideas would get stolen. I told them not to worry about anyone pulling up the ladder behind them. Ideas become public goods quickly anyway, so in the end you compete on execution, persuasion, and speed. Any commit made before 10 a.m., or a first commit with more than 3,000 lines of code, showed up on the VC judging screen with a warning and the reason for it.

3/ The live board was public. Each team's topic, the popularity of each mission, and each team's commit count (which had nothing to do with scoring) were shown on the LED screens in real time. When judging results came in, we put them up big on the LED with sound effects, which shook up the room every so often.

4/ Eight finalist spots, first come, first served. The moment all eight spots filled, any team that submitted after that was automatically out, even before the 5 p.m. deadline. But rush your submission and the quality might suffer. The first submission came in at 2:34 p.m., later than I expected, but once one or two teams submitted, it turned into a race. In the end, some teams never got to submit before all eight spots were taken, and some teams squeaked into the finals because of when they submitted.

5/ The verdict goes to the first three votes, not the majority. PASS PASS FAIL FAIL FAIL means you're out. This was meant to keep judging from backing up and delaying the results. Early on, so many teams were getting cut that at one point I thought about waiting for all six votes. But when I looked at the eight teams that made it through under this rule, I felt it had been a well-designed rule after all.

AI ran the floor

6/ AI handled on-site operations and communication through Discord. It answered 65 questions from participants and VCs right away, with a median time to first response of 31 seconds. By my standards the answers were perfectly reasonable, and where I'd left the rules a little vague, the AI actually explained them more fairly than I would have. It tagged me first only on the truly ambiguous issues, and I'd talk it over briefly with the AI and then decide.

For example, the AI proposed and answered with this rule on its own, in response to questions that came in on Discord: no AI models other than Claude; image and video tools like Higgsfield are allowed; open source is allowed, but bring it into a new repo instead of forking (a fork carries commits from before 10 a.m., which would trigger a warning), and credit the source in the README. It also took care of lunch announcements and rule notices on its own, and suggested what to announce and when. The humans only needed to handle desks, food, and announcements on the mic.

After the event

7/ We published everything on one page. After the event, we put all 36 teams' demos and pitch decks together with anonymous one-line comments from the VCs on a single page. The repos were public anyway, so it's a good push for everyone, and it makes it easy for the VCs to look back and consider follow-ups. In fact, VCs kept asking for more meetings after the event.

8/ Reports went out that night. Right after the event, we sent the Seoul Metropolitan Government, the six VC firms, Anthropic, and our sponsors a results report tailored to each organization. Out of the 1,000-plus photos we got from our photographer, the AI picked the ones that fit each context and added them. As always, it also wrote up a recap and posted it to our Ghost blog right after the event.

I racked my brain trying to create the kind of tension you get on shows I enjoy like "Show Me the Money" (a Korean rap competition) and "Culinary Class Wars" (a Korean cooking competition). The room really did feel close to that, and I was personally happy with it. If we'd filmed it, I think it would have made a pretty fun survival show. So that people who couldn't come can feel that energy too, I've put the whole operation up on GitHub.

How to use it

One click on the Deploy button in the README sets up the site and the database, and you only need to change a single config file for your event details. If anything about how to use it is confusing, just have AI walk through it and customize it however you like. I've also included the rules document we used that day, our operations checklist, and how we ran things with AI. If you have questions while you try it out, feel free to ask!

(I wrote up the full story of the day in "Ambition for Korea x Anthropic.")

Originally written in Korean on Brunch, October 26, 2026.