I watched Erik Schluntz's talk Vibe Coding in Prod three times. The first time for the content. The second time to see what I had missed. The third time because I realised the talk was not really about coding.
Schluntz is a researcher at Anthropic who broke his hand biking to work. He spent two months in a cast. During that time Claude wrote all of his code. His team recently shipped a twenty-two-thousand-line change to their reinforcement-learning codebase. The vast majority of it was written by the model.
The line that stopped me:
You are the product manager of Claude, not the user of Claude. Ask what you can do for Claude, not what Claude can do for you.
The flip is the whole insight. If you treat the model as a tool, you get tool-grade output. If you treat the model as a capable contractor who needs a crisp brief, context, acceptance criteria, and a chance to ask clarifying questions, you get something closer to what a senior engineer would ship.
Schluntz spends fifteen to twenty minutes planning with Claude before a single line of code gets written. The model reads the codebase, finds relevant files, asks about constraints, flags the edge cases. Only then does he fold all of it into a single prompt and say, "now build."
I had been doing it backwards. Typing a vague ask, watching Claude guess, correcting it, guessing again. The AI equivalent of reviewing a junior developer's code without telling them what the sprint goal was.
The second I adopted the PM posture, two things happened. First, my output quality went up. Second, I stopped being the bottleneck. The model was never the limiting factor. My brief was.
The shift: I had been advising clients that AI is a collaborator, not a vending machine. I was still half-treating my own tools like the vending machine.