Claude needs a contrarian.

The model isn't the problem. I'm the problem. I trust it. What I needed was a friend who doesn't — so I cloned one.

Eight days into building my stack on Claude, I looked back at the failures. Most of them weren't model failures. They were me failures.

Claude offered an answer, I took it. Claude proposed a plan, I ran it. Claude suggested a fix, I shipped it. As Andrej Karpathy put it on X in January: "The models make wrong assumptions on your behalf and just run along with them without checking."

I was the wrong-assumption catcher. I wasn't catching enough of them.

Kendall Roy — corporate fintech, big-corporate scars — is the most usefully contrarian person I know. You tell him a great idea, he sees the three things you've missed. You tell him a great plan, he laughs in your face and tells you why it won't work. He finds the one card in your house that won't hold and he flicks it.

Built an agent of him on Tuesday afternoon. Not the real Kendall. A persona Claude can read before it answers me. Sits between me and the model. Before any plan, any deploy, any "here's what we'll do," Kendall reads it first. Pushes back. Asks for four more options. Says it's too complicated. Asks where the rollback is. Checks whether Claude had the context to give a good answer in the first place.

He's reviewed three projects since Tuesday. Each time, the answer that comes back is sharper than what would have shipped without him.

The funny part: Kendall (the real one) sent me a message yesterday. He'd been thinking, independently, about how he was effectively the human version of the agent I'd just built. He didn't know I'd cloned him. He recognised himself.

What I install for clients now: the contrarian inside the workflow. Most operators trust Claude the way I did — they nod, they take the answer, they ship. The fix isn't a better model. It's a second voice that starts at "no" and makes every plan earn its way up. The same pattern runs the rest of my AI team too.

Quite funny, really.