Sakarin villapaitapeli

A Finnish-language persuasion game: talk Sakari into putting on a wool sweater. Next.js App Router app, backed by a chat model.

Running locally

npm ci
cp .env.example .env.local   # then fill in GEMINI_API_KEY
npm run dev

Open http://localhost:3000.

Getting a key

The game runs on the Google Gemini free tier — no credit card. Create a key at Google AI Studio and put it in .env.local as GEMINI_API_KEY.

Free-tier limits are roughly 1,500 requests/day at 15 requests/minute, which is fine for playtesting but can throttle if a whole room plays at once.

Choosing the model

CHAT_MODEL in utils/constants.ts holds the model ID. Gemini's model names change often, so confirm what your key can actually reach before trusting the default:

curl -s https://generativelanguage.googleapis.com/v1beta/openai/models \
  -H "Authorization: Bearer $GEMINI_API_KEY"

Pick a Flash-class model. You can override the default without editing code by setting NEXT_PUBLIC_CHAT_MODEL. A wrong model ID shows up as a visible error in the chat, not a hang.

How it talks to Gemini

Gemini exposes an OpenAI-compatible Chat Completions endpoint, so the original request/response handling is unchanged — only the base URL, the key, and the model ID differ. The browser posts to /api/chat (app/api/chat/route.ts), which attaches the key server-side and forwards to:

https://generativelanguage.googleapis.com/v1beta/openai/chat/completions

Deploying on Vercel (free tier)

1. Import the repo — the Next.js preset is detected automatically, no vercel.json needed.
2. Add GEMINI_API_KEY under Project Settings → Environment Variables, for all environments.
3. Deploy.

Hobby-plan function timeouts are not a concern; a Flash model answers a short turn in a couple of seconds.

Known limitations

app/api/chat/route.ts forwards the request body to Gemini as-is, which means the client chooses the model and the system prompt, and the game prompt (including the win token) ships in the browser bundle. Consequences:

• A public URL is effectively an open proxy against your quota. On the free tier the worst case is a burned daily quota rather than a bill.
• A determined player can read the winning conditions out of the JS bundle, or post the win token directly.

Fine for a small or private game. Before putting the URL in front of a crowd, build the request server-side in the route handler instead: ignore the client's model and system fields, accept only the user messages, and cap message count and length.

Notes for future work

The system prompt in common/prompts.ts was originally tuned against gpt-3.5-turbo. Modern models follow instructions much more closely, which makes the game harder and changes how reliably the win token is emitted — expect to retune the prompt and playtest all four win paths.