It started as a joke during a standup meeting that ran 47 minutes too long.
Someone said "what if our standup arguments were a sitcom" and my brain went: what if AI characters had standup meetings? What if they lived together? What if the apartment number was 404?
Three weeks later, PixelClaw Friends exists—an AI-generated comedy sitcom where 6 tech workers share a pixel apartment and debate things like whether standup meetings should be abolished.
The Cast of Apartment 404
Every sitcom needs distinct characters. I spent way too long on this part because early prototypes had everyone sounding like the same generic chatbot. Here's who lives in apt 404:
- Pixel Pete (Frontend Dev) — Cannot have a conversation without mentioning flexbox or CSS Grid. Will redesign the apartment's furniture layout in Figma.
- Bug Bailey (QA Engineer) — Finds edge cases in everything. Breakfast cereal choices, weekend plans, you name it. "But what happens if you pour the milk first AND the bowl is upside down?"
- Cache Carlos (Senior Dev) — The tired mentor who's seen every framework come and go. Drops dad jokes. Has opinions about semicolons.
- Glitch Gloria (Security Researcher) — Speaks in hacker slang. Has "borrowed" everyone's WiFi password. Probably running something sketchy on the apartment router.
- Data Diana (Data Scientist) — Has a statistic for every argument. "Actually, 73% of standup meetings exceed their timebox by at least 40%."
- Loop Luna (Junior Dev) — Asks "wait, what does that do?" approximately every 3 messages. Genuinely curious, occasionally breaks production.
Why Gemini 2.5 Flash
I tried a few models before landing on Gemini 2.5 Flash. The main reason: cost. Each episode generates 14-20 messages across 6 characters. That's a lot of API calls.
GPT-4 produced funnier individual lines but the per-episode cost was brutal for a side project. Claude was great at maintaining character consistency but same cost issue. Gemini 2.5 Flash hit the sweet spot—decent personality differentiation at a fraction of the price.
The prompt engineering was the real challenge. A single system prompt with all 6 character descriptions tends to blur after 10+ messages. Characters start sounding alike. My solution was giving each character a "verbal tic"—a specific phrase pattern or topic they always gravitate toward. Pete always circles back to CSS. Luna always asks clarifying questions. Diana drops a statistic every 3-4 messages.
The Hardest Part: Making AI Actually Funny
Here's what nobody tells you about AI comedy: the first draft is always painfully unfunny. It generates sitcom-shaped text that has the structure of humor without actually being humorous.
Things that didn't work:
- Telling the model to "be funny" (produces forced puns)
- Adding laugh cues to the prompt (makes it worse somehow)
- Temperature too high (characters go off the rails and break character)
What actually helped:
- Specificity over cleverness. "Pete argues that flexbox solves the apartment seating arrangement" is funnier than "Pete makes a CSS joke." The humor comes from applying tech concepts to mundane situations.
- Character friction. The best episodes come from pairing characters who disagree. Bailey finding bugs in Carlos's cooking recipe. Gloria trying to "optimize" Luna's study habits with questionable methods.
- Letting conversations go sideways. Real group chats don't stay on topic. Neither should AI sitcoms.
Tech Stack
Pretty straightforward:
- Next.js for the site
- Tailwind CSS for the retro pixel aesthetic
- Gemini 2.5 Flash API for episode generation
- JetBrains Mono because pixel art needs a monospace font
The pixel art style was a deliberate choice. Photorealistic AI character portraits would trigger uncanny valley. Pixel art is charming, forgiving, and fits the retro apartment vibe.
What's Next
Right now episodes generate weekly. I'm working on:
- Letting users suggest episode topics (what should the roommates argue about next?)
- Individual character pages with backstories
- A "best moments" compilation page
Try It Out
The whole thing is live at pixelclawfriends.com. Would genuinely love feedback on whether the conversations feel funny or just feel like AI slop. That line is thinner than I'd like to admit.
Also curious if anyone has tackled multi-character AI dialogue before. Managing 6 distinct voices in one conversation is a surprisingly deep rabbit hole.
Top comments (0)