Claude Clue #9: Let Your Models Fight Each Other

One of the most useful things I've found is putting models in adversarial roles against each other, or even against themselves in different contexts. The idea is simple: have one model create something, then ask another model (or the same model with fresh context) to tear it apart. No shared history, no blind spots from the conversation that built the thing. You can target security issues, API interactions, efficiency problems, or just ask it to find anything that seems off.

Here's what this looks like in practice: when I finish a coding session with Claude, I'll ask it to generate an adversarial prompt specifically designed to critique what we just built. Then I take that prompt to a fresh Claude instance (or a different model entirely) and let it loose. The fresh perspective catches things we both missed because we were too deep in the weeds. I've used this same approach with Survey123 data collection—using the built-in models to gather info, then running an external model as QC. It's like having a code reviewer who wasn't in any of the meetings and has no reason to be polite about your choices.

Claude helped write this. We're not hiding it—that's the whole point.

Thanks to Tim Rawson for helping put these together every week!