Sota Misaki Writing

Running the Gemini-1 Grand Prix

September 18, 2026

From July 2025 to February 2026, the N High Group student council and Digital Committee ran Gemini-1 Grand Prix, a recurring contest where students used Google’s Gemini to explore how much the way you prompt — and the way you think about a problem — can change the result.

I wrote the official N High Group blog post about the project as well.

That article covers what the contest was and what students created. This post is the other side of the story: operating it, watching submissions fall, trying to understand why, and changing the system while it was still running.

The problem we kept hearing

Generative AI was already familiar to most students. Knowing that Gemini existed, however, was very different from knowing what to do with it.

We repeatedly heard variations of the same two problems:

There was another problem too.

Using AI could easily become the goal in itself: write a vague prompt, accept a vague response, and call the process finished. That was not the kind of AI use we wanted to encourage.

So instead of running another workshop that ended with “try Gemini sometime,” we built something students could enter repeatedly.

Participants designed their own prompts, submitted the resulting output, and explained what they were trying to achieve. The process behind the output mattered just as much as the output itself.

Six rounds, 3,411 submissions

We ran six rounds between July 2025 and February 2026.

Students could submit more than once because iteration was part of the point. We did not want the contest to reward only one polished attempt; we wanted students to experiment, improve their prompts, and try Gemini in different situations.

Submissions were evaluated across four areas:

By the end of the project, we had received 3,411 submissions.

But the total alone hides the most interesting part of the story.

Use cases submitted by round

Lighter bars: mostly 3rd-term ops. Darker bars: 4th-term ops (Nov–Feb).

The July–August round started with 336 submissions.

Then participation dropped sharply:

The final February round received 2,822 submissions.

The early rounds were primarily operated by the third-term student council. From November onward, the fourth-term student council took over operations, and I was involved as the Digital Committee representative.

For us, the challenge was no longer simply “run an AI contest.”

It was: why are students not submitting, and what can we change about the system?

The real problem wasn’t Gemini

When we took over, the hardest problem was not Gemini itself, and it was not the judging rubric.

It was participation.

A contest can have a good concept, good prizes, and good judging criteria, but none of that matters very much if entering it feels inconvenient or unnecessary.

I started treating the low submission count as an operational problem and looking for friction in the participation flow.

Several things stood out.

1. Public posting created unnecessary friction

Originally, submitting an entry involved running a Slack workflow and posting both the prompt and the generated result into a channel.

That meant participation was also publication.

For some students, that was fine. For others, it meant showing unfinished ideas, experimental prompts, or work they simply did not want to broadcast to a large school Slack workspace.

Those are two different decisions:

  1. I want to enter the contest.
  2. I want everyone to see what I submitted.

We had effectively bundled them together.

So we separated them.

Posting a use case publicly remained possible, but publishing it to the channel was no longer required in order to participate.

It was a small rule change, but it reflected a broader principle I came away with: the minimum action required to participate should actually be the minimum action required to participate.

Anything else is friction.

2. Make repeat submissions matter

The contest included judged awards, but we also wanted repeated experimentation to have value.

For the lottery component, we introduced a weighting coefficient based on the number of submissions.

Every 10 submissions increased the lottery weighting coefficient by 10%.

For example:

The intention was not to make the first entry less valuable. It was to give students a concrete reason to keep experimenting after their first submission.

“Submit again” works much better when another submission represents visible progress rather than simply more work.

3. Make the leaderboard feel reachable

We also changed how we talked about low participation.

Instead of hiding weak submission numbers behind generic announcements, we used the situation itself as part of the message.

If relatively few submissions had been made, then the leaderboard was still reachable. That was useful information for someone deciding whether entering was worth their time.

We also contacted the current top 10 participants directly, told them where they currently stood, reminded them that the ranking could still change, and encouraged them to keep submitting — whether to pursue the highest submission count or to improve their lottery weighting.

The goal was to make the state of the contest visible.

A leaderboard only motivates people when they can understand where they are relative to everyone else.

4. Separate operations from participation

Another issue was where we were running the contest.

A large part of the operation lived inside the main school Slack workspace. That made operational conversations easy to bury among ordinary school activity and blurred the line between the people running the contest and the people participating in it.

We connected the student council Slack workspace and the school workspace using Slack Connect.

That let us keep more of the operational side within the environment used by the student council while still connecting the project to the space where students were already active.

Technically, it was an infrastructure change.

For me, though, the bigger improvement was organizational: it became easier to separate running the event from participating in the event.

Then February happened

February ended with 2,822 submissions.

That is roughly 83% of all 3,411 submissions made across the six rounds.

It is tempting to look at that graph and credit one change.

I don’t think the data supports doing that.

We changed several things around the same period: the participation flow, privacy expectations, incentives, outreach, and parts of the operational structure. We did not run a controlled experiment, so I cannot tell you that any single change caused the February spike.

What I can say is that we stopped treating participation as something that would happen automatically.

We actively worked on the reasons someone might decide not to submit.

Some of those reasons were about friction. Some were about incentives. Some were simply about whether students remembered the contest existed and felt that entering again was worthwhile.

The graph changed dramatically after that work.

That does not prove causation, but it made one lesson very clear to me: participation itself has to be designed.

What I took away

The hardest part of Gemini-1 Grand Prix turned out not to be prompt engineering.

It was product and community design.

If someone has never seriously used Gemini before, the first interaction has to feel approachable.

If someone has already entered once, there needs to be a reason to experiment again.

If submitting something requires making it public, you should ask whether publicity is actually necessary.

And when participation is falling, rewriting the same announcement with slightly different wording is probably not enough. You have to look at the system around the announcement.

Gemini-1 Grand Prix was also a useful reminder that an educational project can be treated like a product.

You can look for friction.

You can redesign the user flow.

You can change incentives.

You can observe what happens.

And, just as importantly, you can be careful not to claim that a correlation proves which change worked.

The contest ultimately received 3,411 submissions, including 2,822 in the final round. More important to me than the final number, though, was learning how much the design around an activity affects whether people actually participate in it.

The school-facing version of the project is available on the N High Group blog.

If I were building something similar again, I would keep three ideas from the start: make the first contribution easy, make repeated contributions meaningful, and design for the person who is interested enough to participate but is still one small piece of friction away from giving up.