Building PartAI: What We Learned Creating an AI Party Game

Two years ago, I was at a friend's game night playing Jackbox when someone's prompt to draw "Coffee" resulted in what looked like a brown alien having an existential crisis. We all laughed, but I remember thinking: what if instead of terrible drawings, we used AI? Not as a gimmick, but as an actual game mechanic that created genuinely interesting moments?

That random thought turned into PartAI. And holy hell, building it taught us things about AI, gaming, and human creativity that we definitely didn't expect.

The Original Concept (Which We Immediately Broke)

Version 0.1 was embarrassingly simple: players got a word, wrote a prompt, AI generated an image, others guessed the word. Classic Pictionary rules with AI instead of drawing.

We thought we'd launch in three months.

Six months later, we were still debugging why the AI kept interpreting "Apple" (the fruit) as "Apple" (the tech company with suspiciously phallic headquarters). Turns out prompt engineering for a party game where players are intentionally trying to break your system is... different from normal AI work.

Problem 1: People Are Creative in Chaotic Ways

Early playtests revealed something we should have anticipated: give people an AI image generator in a competitive party game setting, and they will immediately try to generate the most cursed, unhinged content imaginable.

Not maliciously. Just... creatively. Someone got the word "Doctor" and wrote "person wearing white coat covered in red liquid holding sharp instruments while screaming." Technically accurate. Also absolutely horrifying.

We had to build content filters. But not overly aggressive ones—we didn't want to kill creativity. Finding that balance took months of tweaking and analyzing thousands of real game prompts.

The Tech Stack (And Why We Chose It)

People always ask about our tech choices. Here's the honest answer: we picked tools we knew would work, not tools that were trendy.

Next.js + Supabase: The Pragmatic Foundation

We needed real-time multiplayer. We needed it to scale without exploding our budget. We needed to ship fast because we're a small team and rent exists.

Next.js gave us server-side rendering for SEO and fast initial loads. Supabase gave us real-time subscriptions, authentication, and a Postgres database with an actually decent admin interface. Could we have built a more "elegant" custom solution? Sure. Would it have taken six extra months? Absolutely.

Best decision we made: using Supabase's real-time features for lobby updates instead of building our own WebSocket infrastructure. Saved us probably hundreds of hours of debugging connection drops and synchronization issues.

The AI Image Generation Dilemma

This was the hardest part. We needed images generated in under 10 seconds. Preferably under 5. Because having players wait while "AI is thinking..." kills the party game energy instantly.

We tried:

• DALL-E 2: Good quality, too slow (15-20 seconds)
• Stable Diffusion 1.5: Fast but quality was hit-or-miss
• Midjourney: Amazing quality, no API at the time
• Stable Diffusion XL: Better quality, still kinda slow

We eventually landed on DALL-E 3 for lobbies that prioritize quality and Flux Schnell for lobbies that prioritize speed. Let users choose their own tradeoff. Seems obvious in retrospect, but we spent way too long trying to find "the perfect model" before realizing perfect is different for different groups.

The Game Modes We Almost Didn't Build

Classic mode was the plan. Everything else happened because playtesters kept asking "but what if..."

Human Verification: An Accidental Discovery

During a playtest, someone jokingly wrote a super formal answer to "what did you do this weekend" as a meme. Everyone immediately accused them of being an AI impostor. We weren't even testing that mode yet—it didn't exist.

But the moment happened naturally. Players were already thinking about AI vs human text. So we built Human Verification mode based on that single organic moment. It's now one of our most played modes.

Fake News: The Mode That Scared Us

We debated for weeks whether to include Fake News mode. Is it socially responsible to gamify misinformation creation?

Ultimately, we kept it because it's educational. Playing it teaches media literacy. You learn what makes headlines believable, what red flags to look for, how easily plausible-sounding lies can be constructed. We added an educational disclaimer and made sure the mode emphasizes detection as much as creation.

Player feedback has been overwhelmingly positive. Teachers have literally reached out asking if they can use the mode in classrooms. Which is wild and also kind of heartwarming.

Performance Nightmares and How We Fixed Them

The Great Lobby Cleanup Incident

Three months after launch, our database was absolutely wrecked with thousands of abandoned lobbies. Players would create a game, never actually start it, and the lobby would just... sit there. Forever. Accumulating.

We implemented auto-cleanup: lobbies close automatically after 3 hours of inactivity. Seems obvious but we genuinely didn't think of it initially. Rookie move.

Image Caching Saved Our Budget

Early on, every prompt generated a new image via API call. Expensive. Really expensive. We were going broke during viral spikes.

We implemented aggressive caching: if someone uses the same prompt within 24 hours, serve the cached image. Reduced API costs by about 40%. More importantly, identical prompts now generate identical images, which prevents weird inconsistencies players were noticing.

Community Surprises

We expected competitive gaming crowds to try PartAI. That happened. What we didn't expect:

• Remote work teams using it for icebreakers. Apparently Human Verification mode is great for helping new remote hires get to know their team's communication styles. Cool unintended use case.
• Language learning groups. Non-native English speakers use our game to practice casual English in low-pressure situations. One group plays every Tuesday. They're better regulars than we are.
• Creative writing communities. Unfinished Story mode turned into a collaborative fiction workshop for some Discord servers. They screenshot the best endings and compile them into actual short story collections.

None of this was planned. All of it makes us really happy.

What We'd Do Differently

1. Build moderation tools from day one. We added them later and retroactively cleaning up edge cases was painful.

2. Invest in better onboarding. First-time users are sometimes confused about how to join lobbies. We should have tutorial flows from the start.

3. Mobile design earlier. We built for desktop first, then realized tons of people play on phones in voice chat together. Mobile should have been priority one.

4. More varied AI models sooner. We stuck with one model too long. Offering choices earlier would have made different player groups happier faster.

The Weirdest Bugs We've Encountered

The Penguin Problem: For about two weeks, any prompt involving "Antarctica" or "ice" had a 70% chance of generating a penguin, even if penguins weren't mentioned. We never figured out why. It just... stopped happening eventually.

The Midnight Madness: Users in certain time zones experienced way higher rates of bizarre AI outputs between midnight and 2 AM local time. Turns out our AI provider was doing model updates during those hours and serving slightly different versions. Now we cache more aggressively during those windows.

The Accidental Name Collision: A player named their lobby "test" which collided with our internal test database lobby ID. Oops. Namespacing is important, folks.

What's Next

We're working on:

• Custom word packs (player-created word lists for inside jokes)
• Tournament mode (official PartAI championships because apparently that's a thing people want)
• Better spectator features (for streaming)
• Voice integration (so you don't need Discord open in another tab)

But honestly, the coolest part of building PartAI has been watching what players do with it that we never anticipated. Every weird, creative, chaotic way people use our game teaches us something about how humans interact with AI in social contexts.

This was supposed to be a simple party game experiment. It turned into a weird little community of people who like yelling at their friends about whether a robot wrote their trivia answer.

Couldn't be happier about it.