GDG Basilicata put 35 of 100 hackathon points on the spec the agent reads

The file worth 35 points is the file the agent reads. Teams wrote SPEC.md first, a mentor approved it, and only then did Google Antigravity, Cursor or VS Code with Gemini Code Assist scaffold the app and wire up the Gemini calls ^[3][4][6]. Most rubrics score an artifact nobody in the build loop consumes. Here a vague functional requirement costs a team twice: once when a judge reads the document, and again in the code the agent hands back against it ^[3][6].

Judges scored the spec, the working demo and the pitch, and nothing else ^[8]. The pitch was worth 30 ^[7]. Subtract both and the running software is worth 35 of 100, the same as the document describing it ^[9].

The mentor signature is the enforcement ^[4]. Without that gate the spec gets written after the demo works, and the 35 points measure documentation instead of design. Ninety minutes out of eight hours is about a fifth of the event spent before anyone opens an editor ^[10]. Some teams rewrote their spec twice before a mentor approved it ^[5].

Teams with no prior exposure to writing specs stalled in the first phase, so the quickstart doc now ships a pre-filled file with four headings: Obiettivo, Requisiti Funzionali, Requisiti Non Funzionali, Output Atteso ^[16].

The compute bill is zero. The API key is free and the IDEs are free or already installed ^[20][2]. The cost moved to people and network. Five roles ran the day ^[19]. Mentors were kept off the scoresheet and carried spare API keys for students whose AI Studio redemption stuck ^[8][18]. The organisers asked campus IT for a dedicated SSID in advance and kept 4G hotspots on hand, because the university Wi-Fi struggled under dozens of simultaneous connections ^[17]. On the deadline, the write-up says write access closed automatically at 17:00 with no grace period, and explains why: "the first team you grant one to will tell the next twenty it's negotiable" ^[14][15].

The twenty-minute prototype is the organisers' own claim about the room they ran, and their stated reason for not grading prototype output ^[11]. The write-up does not give team scores or which entries won, so whether the spec weighting spread teams out better than a demo-heavy grid would have is an open question ^[21]. A prototype has to be genuinely cheap in your domain, which it is when the deliverable is a browser app calling a hosted API and the allowed outputs are a web app, a dashboard or an agent ^[13][2]. And you need enough mentors to clear an approval queue inside the first ninety minutes, given that the same mentors are also fixing API access ^[4][8].

The four tracks were study tools for university life, services for local government and territory, autonomous agents, and multimodal apps mixing text, image, audio or code ^[12]. All four run in the browser on the hosted API ^[2].