Your PRD Makes 20 Claims About Customers. You Can Back Maybe 12.

Open any PRD in your org. It's a stack of claims about customers. "Users struggle with onboarding." "Enterprise prioritizes SSO." "The top objection is pricing."

Every one of those is load-bearing. If the claim is wrong, the feature on the other side of it is wrong.

In 2026 the claims got cheaper and the checking got harder

Your PRD is AI-drafted now. You paste the brief, and twelve minutes later you have twenty tidy customer claims arranged into user stories.

Some are real, anchored to something a customer said. Some are pattern-matched extrapolation the model produced because the prompt made them plausible. They read identically.

You won't know which is which until someone calls you on it in review, or six months later in the post-mortem, when the feature underperformed and nobody can reconstruct whether the evidence was ever there.

That isn't a quality-of-PMs problem. Confident sentences about customers are free to produce now. The bottleneck moved to checking them, and that's the half nobody has tooling for.

---

/prd-validation, part one: cross-examine every claim

Paste the PRD. Each claim becomes a query against your conversation corpus. Three verdicts come back, each with the receipts.

"Onboarding friction is the top adoption blocker for mid-market."

Supported
14 quotes, 9 accounts, last 90 days, clustered by sub-theme. Paste the count and the clips inline. The claim has a footnote now, and the review moves on.

"Customers want a simpler dashboard."

Contradicted
2 asked for simpler. 6 asked for more density and control. Rewrite the line yourself, before the room does it for you.

"Pricing flexibility is the top reason prospects walk away."

Unsupported
Nothing in the corpus. There's plenty on pricing, but the top objection is integration coverage, not flexibility. This is the bucket where the AI extrapolations live, and it's the one that matters most.

The most useful sentence a check can return is "nobody said this." It's the one you can't get from the model that wrote the claim, because it's the model's blind spot by design.

---

Part two: what the PRD never mentions

Cross-examination grades what's on the page. The bigger win is what isn't.

/prd-validation reads the corpus for the area your PRD covers and surfaces what you left out. Say you're scoping a notifications redesign. The claims hold up. Then it hands you the rest:

• 7 accounts asked for batching or digests. Not in the PRD.
• 4 finance accounts need an audit trail on delivery. Compliance edge case. Not in the PRD.
• 3 enterprise accounts asked for per-channel routing. Not in the PRD.
• 5 flagged the current system breaking under load. A failure mode your happy path ignores.

This is the move every PM knows they should make and never has time for. "Let me re-watch fifty hours of calls to be sure I didn't miss anything" does not happen. Here it takes a minute, because the corpus is already typed, attributed, and searchable.

---

What it changes about the review

The PRD review you know has a rhythm. Someone challenges a claim. You either pull up the receipt or fall back on "that came up in a few calls." The decision gets made on who pushed back hardest, not on the evidence.

When every load-bearing claim is tagged Supported, Contradicted, or Unsupported, and the doc has already absorbed what the corpus surfaced, the meeting changes. You stop defending and start presenting. The argument is no longer "do we believe this." It's "given this, what do we build."

And nobody comes back in two weeks with "we forgot about X," because X was in the corpus, and the corpus was in the room.

A PRD without evidence is a hypothesis. Run it through Fragment first, and it's a decision.

Paste a PRD into /prd-validation and see which claims survive.