Writing PRDs when AI drafts them for free

The document was never the point

Anyone can generate a passable PRD in about 60 seconds, which means the PM’s value has moved entirely to the judgment the document used to hide. Feed ChatPRD or a raw model a one-line prompt and it returns a clean structure: problem statement, user stories, success metrics, edge cases, a rollout plan. The output is better than what most PMs wrote by hand three years ago. So the takes landed where they always do: the PRD is dead, the PM is next.

That reads the situation backwards. The PRD isn’t obsolete. The parts of the PRD that were always clerical got automated, and what’s left is the part that was always the actual work: deciding what problem is worth a team’s quarter. The document became a commodity. The thinking behind it is now the entire job, exposed with nowhere to hide.

A PRD did two jobs, and only one just got automated

A PRD did two jobs, and they were never equally valued.

The visible job was documentation, a shared artifact so engineering, design, data, and leadership all worked from the same spec. This is the job everyone talks about, and it’s the job AI now does for free. Structure, formatting, exhaustive edge-case enumeration, acceptance criteria phrased so nobody can claim they weren’t told: all of it is pattern-matching over a corpus of a million prior PRDs. A model is better at that than you are, faster, and it never gets bored on the fourth edge case.

The invisible job was forcing a decision. Writing a PRD made you commit to a problem, a user, a definition of done, a set of things you were explicitly not building. The blank page was a filter. If you couldn’t articulate the problem crisply, the document stalled, and the stall was information. It told you the thinking wasn’t finished. Plenty of bad projects died in the drafting because the PM couldn’t make the argument hold together on paper.

Most PMs experienced these two jobs as one activity, so they never separated them. AI separated them for us. It took the visible job and left the invisible one sitting on the table.

Free drafting exposes weak thinking instead of hiding it

Here’s what a team learns the first month it adopts AI drafting. The PRDs get longer, cleaner, and more numerous. And the quality of decisions doesn’t move at all.

That’s the tell. If the document were the constraint, better documents would produce better products. They don’t, because the document was never the constraint. A beautifully specced feature solving a problem nobody has is still a beautifully specced waste of a quarter. AI drafting removes the friction that used to slow a bad idea down, which means bad ideas now reach engineering faster and dressed better.

The stall I described, the blank page that wouldn’t fill because the thinking wasn’t done, is gone. A PM who hasn’t decided what problem they’re solving can now generate 2,000 words that look like a decision. The model will invent plausible user stories and success metrics for a problem the PM hasn’t validated. The artifact is confident. The thinking underneath is empty. And the org can’t see the difference from the document alone, because the document is no longer evidence that anyone thought.

The floor for producing a document collapsed. The floor for producing a good decision didn’t move. The gap between those two floors is where weak PMs now hide, and where strong ones separate.

The new template is mostly deletion

If the model writes the spec, the PM’s document changes shape. Most of what used to fill a PRD should now be a one-line prompt and a generated appendix nobody reads unless they’re implementing.

What stays on the human page is short and hard to fake. The problem, stated so specifically it excludes the adjacent problems it’s often confused with. The evidence the problem is real, the ticket volume, the drop-off, the interview quote, the revenue at stake, not a hypothesis dressed as a finding. The explicit list of what you’re choosing not to solve and why, which is the sentence a model won’t write for you because it requires knowing your team’s actual constraints this quarter. And the definition of success that you’d stake your credibility on, not the metric that’s easiest to move.

Four things. None of them are structure. All of them are judgment. The generated PRD becomes the appendix; the human PRD becomes the argument for why the appendix deserves to exist. A PM who can write those four paragraphs well doesn’t need the tool to look competent, and a PM who can’t won’t be saved by it.

The objection here is fair: doesn’t the model help with problem framing too? Ask it to critique your problem statement and it will surface angles you missed. True, and worth doing. But the model critiques the framing you give it. It can pressure-test a decision. It cannot decide which problem, out of the dozen your users would all plausibly pay to solve, is the one this team should spend the next quarter on. That choice depends on your read of the market, the roadmap, the org’s tolerance for risk, and a hundred things that live in your head and nowhere in the corpus.

Problem framing is the one skill that didn’t compress

Every part of product management that could be reduced to a pattern is getting compressed. Drafting, first. Then competitive teardowns, release notes, meeting synthesis, the first-pass roadmap. If your value was producing those artifacts, the value is gone, and pretending otherwise is how careers end quietly. The PM who ships a working prototype in an interview already wins on this axis, because the scarce move is closing the gap between hypothesis and evidence.

Problem framing didn’t compress, and it won’t, because it isn’t a pattern over prior text. It’s a judgment about a specific team, a specific market, and a specific moment, made under uncertainty with incomplete information and consequences you own. The model has read every PRD ever written. It has never sat in your standup, watched your best engineer burn out on the wrong project, or felt the cost of a quarter spent solving the problem that was easy to spec instead of the one that mattered. Value migrates to the scarce input as the output commoditizes.

For a decade the PRD let average PMs look like good ones, because the document did the signaling and the thinking hid behind it. That cover is gone. What’s left on the page is the decision, and the decision has your name on it.