I Stopped Writing PRDs

Reading Time: 3 minutes
The death of the PRD vs the living spec — traditional PRDs versus a living specification verified by AI agents

For most of my career, my job as a product manager started the exact same way: write the document. The PRD was where the thinking happened. And then, if we’re honest, it was where the thinking went to die. Engineering read it once. Design skimmed it. By the third sprint, the document and the actual product had quietly divorced.

I’m building a product again — an AI-agent-run hotel marketplace — and this time there is no PRD. There is a living specification instead, and the difference is not cosmetic. A Stanford workshop I took recently on supervising AI coding agents (TECH 43, taught by Vasyl Rakivnenko) sharpened the approach, but the change itself is simple enough to explain.

A spec is the product, not a summary of intent

A traditional PRD describes what you intend to build. A specification is the product itself, written down precisely enough that it can be built, tested, and argued with.

Mine is a living document: 59 product decisions so far, each one numbered, each one paired with the reasoning behind it. When a question comes up — what happens when a hotel misses its deadline, who can see whose data, when does a price stop being negotiable — the answer isn’t buried in someone’s memory or lost in a chat thread. It is an explicit decision with a number. And if an assumption turns out wrong, we update the decision, and everything downstream reflects it.

The bottleneck moved to supervision

The second shift is who writes the code. I direct AI coding agents against the spec. On the surface, that sounds like a small change in tooling. In reality, it changes what the product manager’s job is.

When the builder can produce code faster than a human can review it, the bottleneck moves — from execution speed to instruction quality and supervision. So the process is built around alignment before execution and verification after:

  • Alignment before code: agents align on what a feature must do — and explicitly what it must not do — before writing a line.
  • Verification as “done”: a feature is not finished when the code runs. It is finished when verification and evals prove it behaves the way the spec decided.

AI doesn’t replace product judgment — it exposes it

People ask whether AI coding agents make the product manager less necessary. So far, I’ve found the opposite.

A long PRD lets you hide behind polite ambiguity. A spec forces every vague instinct into an explicit decision, in writing, where it can be challenged — including by me.

“Qualified = final” is four words. Getting there took a real argument about marketplace incentives: would hotels hold back their best rates if they knew they could keep revising them later? The PRD version of me would have written “hotels can update prices” and moved on. The spec forced me to decide, and that decision made the marketplace mechanics stronger.

What is still unproven

I won’t pretend this is fully proven. The product is not live yet, and a specification is not a business. Real users and real edge cases will test decisions that look clean on paper.

But as a way of developing products, I don’t think I can go back. For twenty years we talked about the product document as the “source of truth.” With AI agents building and verifying directly against a living spec, it finally is.

Leave a Reply

Scroll to Top