The Hypothesis
Party games have a shelf life. Once you've seen all the prompts in Cards Against Humanity, the magic is gone. We hypothesized that AI-generated content could solve the replayability problem — every session genuinely unique, not just reshuffled.
We gave ourselves 14 days and a clear kill condition: voluntary replay rate above 50%.
What We Built
A browser-based multiplayer party game. No app install — scan a QR code, join from your phone. The core stack:
- Next.js for the frontend with real-time WebSocket connections via Socket.IO
- Claude API generating prompts, judging answers, and narrating rounds
- ElevenLabs for text-to-speech — the AI "host" actually speaks
The minimum viable game was AI Quiplash: players get AI-generated prompts, submit answers, and an AI narrator judges and commentates with callbacks to earlier rounds.
The Numbers
Eight playtest sessions, different groups each time:
| Metric | Target | Actual |
|---|---|---|
| Replay rate | >50% | 100% |
| Avg session length | >15 min | 34 min |
| NPS | >40 | 72 |
Every single tester wanted to play again immediately. The kill conditions weren't just cleared — they were obliterated.
The Surprise
Here's what we didn't expect: players didn't care about the AI-generated prompts as much as we thought they would. The prompts were fine. Better than static prompts, sure. But they weren't the hook.
The hook was the AI narrator.
When the AI referenced something a player said three rounds ago — "Ah, classic Sarah, channeling the same energy as your answer about the haunted IKEA" — the room erupted. When the narrator created running jokes across rounds, players leaned in. When it roasted someone's answer with context from earlier, people lost it.
The hypothesis was "AI content solves replayability." The real insight was more specific: AI as performer beats AI as content generator. The value isn't in producing more prompts. It's in creating a dynamic host personality that makes each session feel like a unique live event.
Five Things We Learned
1. Latency Kills Energy
Any pause longer than 3 seconds for AI generation broke the room's energy. Party games run on momentum — the moment someone says "is it loading?" the spell is broken.
We had to heavily optimize: pre-generating where possible, streaming responses, parallel processing of narration and game logic. The technical work of making AI feel instant was harder than making it feel smart.
2. Voice Changes Everything
Our first version used text-only narration. It scored significantly lower in testing. Adding ElevenLabs voice synthesis transformed the experience — suddenly the AI narrator felt like a real host, not a chatbot.
The takeaway: modality matters more than model quality. A mediocre AI response delivered via voice was more engaging than a brilliant response displayed as text.
3. The "Callback" Mechanic Is the Core Loop
The moment that reliably produced the biggest reactions was when the AI narrator referenced something from earlier in the session. This created a sense of continuity and personality that static games can't match.
This isn't just a nice feature — it is the product. Everything else (prompts, scoring, game modes) exists to create opportunities for callbacks.
4. Four Players Is the Sweet Spot
We tested with groups from 3 to 8. The best experience was consistently 4-5 players. Enough for variety in answers, small enough that the AI narrator could track everyone and create personalized callbacks.
5. No Accounts, No Friction
Zero-install browser play via QR code was essential. Every added step (create account, download app, enter code) would have killed the casual "let's play right now" energy that makes party games work.
What's Next
localhost-party cleared all kill conditions and graduated from experiment to active project. The v2 roadmap:
- Multiple game modes beyond Quiplash-style
- Persistent narrator memory across sessions with the same group
- Custom narrator personalities
- Public game rooms for strangers
The core insight — AI as live performer — has legs far beyond party games. But we're staying focused on the game for now. One experiment at a time.
This post is part of our practice of building in public. Every experiment gets a case study — wins and losses alike.