Agents Reading Agents
The artifacts survive. The judgment doesn’t transfer.
The plan was fifty pages. I remember thinking: this is either very good or very wrong, and I have no way to tell which.
I had written the PRD with Claude Code. The engineering manager had taken that PRD and run it through Claude Code to produce the technical plan. Tickets, dependencies, sequencing. Fifty pages of forward motion, all of it derived from a document derived from a conversation with an AI I’m still learning to work with, about a product I thought I understood completely.
When we use AI to boost our productivity we tell ourselves the ‘outputs are what matter’. We tell ourselves this because the outputs are what we can see.
Here is what I know about the PRD I wrote. I know the scope decisions encoded in it were mine, in the way that decisions made in conversation with an AI are yours: you steered, you approved, you moved on. I know some of the assumptions about the user and workflow were ones I had been carrying for months without examining. I know the document did not know any of this about itself.
The engineering manager’s Claude Code did not know it either. It read what I had produced and built faithfully on top of it. Fifty pages of a structure whose foundation I cannot fully vouch for.
The input was assumptions. Unexamined, dressed as decisions. And Claude treated them with the same fidelity it would treat anything else.
The tickets will be groomed. They will be assigned. Someone will pick them up in a sprint and ask whether the approach is right, whether the estimate holds. They will not ask whether the thing should be built at all. That question was settled in the PRD. The PRD was settled when the EM signed off. The EM signed off on fifty pages generated by Claude Code reading a document generated by Claude Code, and somewhere in that chain the question of whether the premises were sound stopped being anyone’s job to hold.
I have been in enough planning cycles to know that humans reviewed PRDs badly long before any of this. The EM who skimmed a document and nodded is not a new figure. A skim still passes the document through a person who has spent years in the domain, who has opinions about what tends to go wrong, who might notice that the framing is off. The pattern-matching happening now is faster and more thorough and entirely indifferent to what the document got wrong.
We are building practices for a technology we do not yet know how to fully vet by using the technology to build the practices. The outputs look complete. Fifty pages looks like due diligence. The tickets look like a plan.
I think about the assumption buried in paragraph four of my PRD. I think about whether anyone will reach it before we’ve built three sprints on top of it.
The artifacts survive. The judgment doesn’t transfer.