The rule that killed "Just Start Coding" at our Hackathon

GDG Basilicata and the IEEE Student Branch at the University of Basilicata, together with the university's Department of Sciences, booked a room built for fifty to eighty students, split them into teams of three or four, and gave them eight hours to build something with Google's Gemini models. Nobody installed a GPU driver. Nobody touched a local model weight. Every team redeemed a free API key from Google AI Studio and worked entirely in the browser.

That's the first pillar of the format we built, and it's the boring one: cloud-first, zero local setup. It matters because the fastest way to kill momentum in a student hackathon is a broken CUDA install at hour two. Remove the infrastructure problem and you get eight real hours of thinking and building instead of six hours of pip errors.

The interesting part is what we did with the time that freed up.

Any team in that room could ask an agentic coding tool for a working prototype in twenty minutes. That capability is common now, which is exactly why we didn't grade on it. Our scoring grid put 35 of 100 points on a file most hackathons never ask for: SPEC.md.

Before touching an IDE, each team spent an hour and a half writing down what they were building, for whom, and how they'd know it worked — functional requirements, architecture, expected output, non-functional constraints. A mentor had to sign off before the team could move to code. We called this phase Spec-Driven Development, and it changed the room's energy. Teams argued about scope before they argued about syntax. Some rewrote their spec twice before a mentor approved it.

Only after that did agentic coding tools come into play — Google Antigravity, Cursor, VS Code with Gemini Code Assist. The agent reads SPEC.md, scaffolds the app, and helps wire up the Gemini API calls. The team's job shifts from typing code to checking what the agent produced against the spec they wrote an hour earlier. That's a different skill than typing fast, and it's the one we wanted to measure.

In the last hour before submission lock, teams didn't write slides from a blank deck. They uploaded SPEC.md, their code, and their documentation into NotebookLM and used it to generate a three-minute pitch script, an FAQ for the Q&A round, and, for teams who wanted it, an audio overview they could play as part of the demo. Thirty of the hundred points went here: how clearly a team could explain what they built and defend it in two minutes of jury questions.

Four tracks were open on the day: study tools for university life, services for local government and territory, autonomous agents, and multimodal apps mixing text, image, audio or code. Output had to land as a web app, a dashboard, or an AI agent, nothing more exotic than that, on purpose — a constraint that stops a team from disappearing into scope an eight-hour sprint can't survive.

The deadline is not a suggestion. At minus sixty, minus thirty and minus ten minutes, someone walks the room and calls it out loud. Write access to the submission folder closes automatically at 17:00, no five-minute grace period — the first team you grant one to will tell the next twenty it's negotiable.

Running this for a room that size, with a jury on top, needs more than a tech lead. We split it five ways.

Mentors and judges are different people doing different jobs. Mentors unblock teams on specs, API access, or NotebookLM extraction; they don't score anything. Judges score the spec, the working demo, and the pitch, and nothing else.

University Wi-Fi struggled under dozens of simultaneous connections, so we asked IT for a dedicated SSID ahead of time and kept a couple of 4G hotspots on hand as backup. A handful of students got stuck redeeming their AI Studio API key; mentors carried spare keys to unblock them on the spot instead of debugging Google's account flow live in front of a team on the clock. Teams with no prior exposure to writing specs stalled in the first phase, which is why the quickstart doc ships a pre-filled template instead of a blank file:

## Obiettivo
<One sentence: what problem does this solve, for whom>

## Requisiti Funzionali
- ...

## Requisiti Non Funzionali
- ...

## Output Atteso
<Web app / Dashboard / Agent — what does "done" look like>

Handing teams a template doesn't lower the bar — a mentor still has to approve it — it just gets them past the blank-page freeze in the first ten minutes.

None of this needs a compute budget. The API key is free, the IDEs are free or already on student laptops, and NotebookLM runs in a browser tab. What it needs is a room, five people willing to own the roles above, three or four mentors, and a rubric that rewards a good spec and a good pitch as much as a working demo. The competition template we cloned on Kaggle — visibility set to Restricted, rubric swapped in, CSV export at lock time — is reusable as-is by any group running a Google-model hackathon.

The retrospective after this run flagged the usual fixes for next time: earlier mentor briefings, a longer matchmaking window, clearer wording on what counts as "multimodal." None of that touches the four pillars. Cloud-first, spec before code, an agent that reads the spec instead of guessing, and a pitch built from artifacts the team already produced — that combination is what let a computer science student and a humanities student sit at the same table and ship the same prototype.