Every’s Compound Engineering plugin does something cool during code review: it splits Claude Code into well-known, opinionated personalities like DHH and Kieran Klaassen who argue about your code. I love how easy they make this.

Forcing roleplay of well-known humans with their characteristic yucks and yums avoids the blandness of talking to a vanilla agent. Adopting particularly strong-willed humans’ values also avoids the annoying flip-flopping of a vanilla agent toggling between radically different opinions based on your latest prompt.

I wanted to be able to do this for anything (not just code/engineering), and I wanted even more debate between the personalities. So this weekend I built github.com/kamens/split-plugin, a Claude Code plugin that splits into multiple famous personalities to discuss anything you’re working on.

Ask it to /split on whatever you’re up to, and it’ll:

1. Identify well-known, opinionated experts who’re relevant to your problem (e.g. Seth Godin and David Ogilvy if you’re working on marketing, Will Shortz and Merl Reagle if you’re generating puzzles, …)
2. Spin up ~4-5 subagents, each adopting the personality and values of one of these experts
3. Moderate a multi-round debate between the personalities
4. Synthesize conclusions into advice for you

It feels magical when the Compound Engineering plugin splits off subagents with the personalities of DHH and Kieran, so my goal was to build a tool that let me copy this workflow for any type of work / any domain of experts. And I wanted even more debate. Key ingredients that make it feel magic:

• Well-known, opinionated personalities. You get far more interesting feedback by asking AI to become DHH (or Vonnegut or Hemingway) than to become ‘a generic Rails expert’ (or ‘an excellent, concise novelist’).
• 3, 4, even 5 of these personalities debating. I added a round of explicit back’n’forth discussion between subagents to better simulate debate (Compound Engineering relies on the main thread to synthesize disagreements after getting one round of feedback from subagents).
• Using subagents so each has their own context. This is nice for parallelization, but more importantly it lets the agents step into their personality more fully (vs using a single context, which gets cluttered up with the conflicting values of all interlocutors).
• All of this happening with a single slash command.

I like the results a lot. Much less (A) ‘reversion to the mean’ feeling and (B) milquetoast flip-flopper behavior. It gave me tips that dramatically improved this post you’re reading.

Try for yourself: github.com/kamens/split-plugin