Claude mirrors what you tell it, building perfect arguments for any side you present. Nick Perry proposes a workflow: ask Claude to clarify your question, spawn five sub-agents to attack it from different angles, then have a chairman agent synthesize their responses into a single verdict.
Claude is the world's most convincing pick-me. See, when you ask it if your idea is good, it'll build a perfect case for why it is. And when you ask if your idea is bad, it'll build a perfect case for why it's terrible. So you can't actually trust it with real decisions before you make an AI boardroom. Here's how. Step 1, ask Claude what you want to ask it. Step 2, before you send it, tell Claude to spin up 5 different sub-agents that attack your question from 5 different angles, and then hit send. What will then happen is a chairman takes all 5 answers and gives you one verdict, one next step. Save this and send it to a friend, please.
