Weird AI Bots: a library of intentionally offbeat AI character configs

Not every AI project needs to be serious, and Weird AI Bots is gleefully unserious. It's a growing collection of system prompts that define intentionally strange, funny, or absurd AI characters — the kind of personalities that make you laugh out loud when you hear them on a speech-to-speech platform like Play.ai or Vapi. I started building these because I was spending so much time writing carefully tuned, serious system prompts for productivity tools that I wanted a creative outlet, and because I genuinely think there's value in exploring the edges of what LLMs can do with personality and character. The collection has grown from a handful of jokes into something I'm oddly proud of, not because it's useful in any conventional sense, but because some of these characters produce moments of genuinely unexpected brilliance.

What happens when you give AI a truly unhinged persona

Here's the thing that surprised me: giving an AI a completely weird persona doesn't just produce weird output — it sometimes produces unexpectedly insightful output. A character configured as a paranoid conspiracy theorist will occasionally surface genuinely useful pattern-matching observations about technology ecosystems. An AI configured as an overly enthusiastic medieval herald will frame modern technology announcements in ways that strip away the jargon and reveal what's actually being said. There's something about removing the default "helpful assistant" constraints that lets the model access different corners of its training data and combine ideas in ways the default persona never would. Not reliably, of course — most of the output is just funny — but often enough to be thought-provoking.

The censorship discovery

A lot of these configs are tongue-in-cheek and poke fun at the current AI fervor, which means they frequently run into refusals from platforms with strict content policies. This turned out to be educationally valuable: the experience taught me more about the varying levels of content restriction across different LLM providers than any amount of documentation reading. Some models will cheerfully roleplay as a grumpy Victorian butler who insults your code. Others refuse to simulate any character that expresses negativity, even in an obviously comedic context. The gap is enormous and largely undocumented, which feeds back into my broader interest in model bias and censorship. I capture most of these system prompts using speech-to-text (meta, I know), so fair warning: there may be transcription artefacts that need cleaning up. The whole collection is open source under CC-BY-4.0 on GitHub. Feel free to remix, adapt, and deploy your own weird bots.