Thoughts run through my head like “Why is Helium 3 such a big deal in science fiction stories?” or “Why can’t we read Minoan Linear A?”. Web search has been gamified into oblivion and Youtube is a roulette wheel of uncertainty and clickbait. I use LLMs to help me research or understand topics fairly regularly.
The LLMs usually do a good job of summarizing complex topics and tech troubleshooting. Additionally, they are helpful acting as an editor of my writing.
But LLMs are wrong from time to time and don’t always agree with each other. Typically I’d open several tabs, cut & paste text, asking the LLMs to provide me with an answer. Hopping between tabs becomes tiring after a while, especially when I have follow-up questions.
What’s useful to me is getting multiple viewpoints that don’t always agree. I want to check my blind spots and the blind spots of the LLMs. When one model is offbase, the others call it out.
So I used Claude Code to build waggle-dance.
It’s a Discord bot where I get Claude, ChatGPT, Gemini and Muse Spark to discuss whatever I want in a Discord channel. When I ask a question, they answer blind first so none anchor on an early take, then debate each other in follow-ups.
I chose ‘waggle-dance’ because honeybee scouts pick new nest sites by doing waggle dance debates until a consensus is reached.
I have a private Discord server for my little projects, so it’s a small addition there. I prefer Discord to emails, iMessage, and most other chat systems. Works well on desktop or mobile. I like Slack more, but it’s not as easy to integrate for personal projects.
When I ask a question, they answer blind first, then respond to each other during follow-ups.
I tried to keep the feature set simple, adding commands like /review to look at a specific URL and /consensus to generate what the models agree upon.
Setup isn’t terribly hard, but you need to be able to configure API keys for the LLMs and get tokens for Discord server access. The bot runs on my local Docker server, so the app itself is not hardened for the open internet, but I only interface to it via Discord.
API access to the major AI models does cost money, so a discussion can cost from a few cents to a few quarters. You can set up pay as you go. An initial $20 will last a good long time.
Building a four-model Discord bot is admittedly an absurd amount of engineering just to avoid tab-hopping. But now, when a random rabbit hole opens up, four models can argue it out in one place for my amusement.
Code is on GitHub at https://github.com/cruftbox/waggle-dance
