Describe how you would version a PRD as model behaviour changes underneath it.

How to answer this, stage by stage

Seven moves. This question sounds like it wants a process. It's really about the day a document stops matching the thing it describes. Each stage has the words you'd actually say.

Let's learn

Here's what happens when a tool changes on a quiet Monday and the document describing it does not.

Say a procurement team uses a tool that reads every contract coming up for renewal and writes a short brief telling the buyer what range to offer the supplier.

Before this tool existed, a buyer spent about three hours before each renewal digging through three years of invoices and drafting their own opening ask by hand. With the tool, that dropped to about ten minutes. It reads the contract, checks the price history, and drafts a brief with a suggested range.

Here's the part that matters. It isn't that the tool got something wrong. It's that nobody had a rule for what happens the day the tool's own wording changes and the sign-off document doesn't move with it.

Twelve claims went into the PRD the week the tool shipped. Claim four says the tool never states a number as final, only a range, so the buyer keeps room to move in the real conversation. When the underlying model got a routine wording update, meant to make the tool sound more confident, claim four didn't change on paper. Nothing tagged that sentence to a version, so nothing told anyone it might now be wrong.

Twelve claims, same size grid, side by side. The broken count barely moved.

At its worst, this costs more than never writing a brief at all. A document everyone trusts and nobody checks doesn't just fail to help. It tells everyone downstream to stop looking exactly when they should be looking hardest.

What I would leave alone. The claim that Renewio flags a contract renewing within ninety days never needed a second look. It's pulled straight from the contract's own date, not from anything the model writes, so it can't drift just because the model's wording changes.

The lesson. A PRD is not a photo of a product. It's a claim about a moving target, and if nobody writes down which version of the target was checked, the claim quietly stops being true and there's no way to tell when.

Now here is the same thing as a story

Use this version when you have room to let it land, not just list it.

The laptop on Chidera Okonjo's desk has one browser tab that's been pinned since the week the tool launched: the PRD.

Chidera has run trust and enablement at Windermere Components for four years. Hand her any escalation ticket about one of the company's tools and she can tell you inside a minute whether it's a real bug or someone misreading what the tool was built to do.

Windermere buys parts from more than two hundred suppliers. Until eighteen months ago, a buyer prepping for a renewal did all the digging by hand. Then the procurement team got Renewio. It reads a contract, flags it ninety days before it renews, and drafts a brief: the supplier's price history, where Windermere has leverage, and a counter-offer range. A range, on purpose, so the buyer keeps room to move once they're actually talking to the supplier.

Twelve claims went into Renewio's PRD the week it shipped. Claim four says the tool never states a number as final. Chidera signed off on all twelve, and for the first six weeks, every time an escalation ticket came in, even the routine ones, she pulled the live chat transcript and checked it against all twelve claims herself, line by line. It always matched.

So she stopped pulling transcripts. She started answering from memory of what the PRD said, the way you answer a question about a rule you've read so many times you stop thinking you need the page.

Then, on a Monday nobody flagged as different, engineering shipped a small wording update to Renewio's prompt. The goal was reasonable: buyers weren't acting on the suggested ranges enough, so the update made the tool sound more confident. Claim four still read the same on paper. Nobody had tagged that sentence to a version, so nobody had a reason to go check it against the new one.

Eleven days later, a newer buyer named Soraya opened a ticket. Renewio's brief for the ArcMetal Supply renewal read, "We're offering $42,000, final." Not a range. She asked Chidera if that was normal.

Chidera didn't pull the transcript. She didn't think she needed to. Claim four. She told Soraya that wasn't something Renewio did, and closed the ticket.

That afternoon, Soraya forwarded the line to ArcMetal's rep, the way the PRD had taught everyone the tool was supposed to work: trustworthy enough to act on without a second read. ArcMetal held the number. Windermere's own price data said the top of a fair range was $37,000. The deal closed at $42,000, with no room left to talk it back down, because Windermere had already called it final.

I want to say the model got worse. It didn't, not really. It got more confident, which was the whole point of the update. The real problem was that Chidera never had a running number for how much of the PRD still matched the live tool. She had a habit, and the habit had exactly two settings: pull the transcript and check, or trust the page. Six quiet weeks flipped it from the first to the second, and nothing was built to flip it back.

People are switches, not dials

So here's the decision I'd take back. When Renewio's PRD was first written, someone on the review floated tagging each claim to the model and prompt version it was tested against, so a version bump would show which claims needed a fresh look. The team talked about it for ten minutes and shelved it. There was only one version of the model at the time. Writing a version number next to every sentence read like paperwork for a problem that didn't exist yet. That was a fair call the week it was made. It stopped being fair the day engineering shipped a change to claim four and the PRD had no way of saying so.

I'd build that tag. Not a bigger PRD, not a stricter sign-off meeting, just a version tied to every claim and a flag when the version underneath it moves. Run the same eleven days through that design. The ticket Soraya opens shows "claim four, unverified since v2.4, 11 days" before Chidera reads a single word of the transcript. She pulls it in eight minutes, sees the drift, and tells Soraya to hold. The $42,000 line never reaches ArcMetal. The negotiation runs at the real range, and Windermere closes closer to $36,000.

That's the whole difference. One design hands the PRD a finish line and calls the job done. The other hands it a pulse, something that keeps checking itself against the model whether or not anyone remembers to ask.

And the part I'd tell myself, if I could go back: we asked whether the PRD was accurate the week we wrote it. We never asked how anyone would find out the week it stopped being.

FLIPS, for a PRD that outlives its own model version

This question sounds like it wants a process. It's really a Perturbation question: a document built to describe a product keeps insisting the product hasn't moved. FLIPS runs straight down the line, F to S.

FLIPS, in five rows

A small move in the model. A hard snap in who was watching it.

And if you want to be sure it really works, try it somewhere else

An HVAC dispatch tool tells field technicians whether to repair or replace a broken unit. Same question, a completely different product, and a different flip.

F. Corbin Ashworth, dispatch supervisor at Fairwind Mechanical Services, who trains every new technician on Unit Advisor's onboarding PRD before their first week on a truck.
L. New technicians read that PRD once, during onboarding, and never open it again, since nothing about the job ever seems to call for it.
I. A different flip. Technicians don't check the tool's advice less carefully. They stop opening the one document that would tell them the advice changed. No middle setting: the PRD gets read during onboarding, or it doesn't get read again, ever, once the first week ends.
P. The onboarding promise: new hires were told the PRD was the one thing they needed to read to understand Unit Advisor, once, and that promise gave their trust nothing to check back against later.
S. Tag the repair-versus-replace rule to the pricing module's version, and force a fresh look the moment the vendor bumps it. A drift that took two hundred and forty days to surface through an angry customer review gets caught in six, before the next onboarding class ever reads the old numbers.

Flashcards (click a card to flip it)

Eight fixed slots, pulled straight from the answer above.

Check yourself Score: 0 / 0