Write the quality bar section for a customer-facing classification feature.

How to say this section out loud

Seven moves. This is a commit question hiding inside a documentation question, so the position comes first and the exception comes later, not the other way round.

One more thing before the walkthrough moves on: an interviewer asking this wants to see you hold a position and still show judgment about when it stops applying. A quality bar that only ever tightens, never once, isn't a bar. It's a mood.

Let's learn

Anchorwell is a banking app built for small businesses. One of its features, called Payment Guard, watches every outgoing payment, ACH transfers and wires, and holds the ones that look like fraud before the money actually leaves the account.

When Payment Guard first launched, the team set the bar low on purpose. It held about 3 payments out of every 1,000. Nobody wanted a business owner annoyed by their own bank on a Tuesday. That bar caught 54 out of every 100 real fraud attempts. It missed the other 46.

Over one year, 40 fraud attempts got through the old bar. The average business recovered only 9 cents on every dollar taken, because a business account doesn't carry the same protection a personal one does.

The decision that mattered. Write the bar as a floor on fraud caught, not a ceiling on holds. A ceiling asks "how few payments can we get away with holding." A floor asks "how much real fraud are we actually willing to let through," and that's the number that should have been in the document from day one.

The lesson. A quality bar written as "keep friction low" will always drift toward the mistake nobody's watching for. Write it as a number you'd have to defend catching, not a number you'd have to defend holding, and the drift stops.

The morning Riverbank's mower money left for good

Use this one when there's room to feel the twelve weeks it took to find out how little comes back. The short version above has the same shape, just none of the Tuesday.

Abena Sarpong is the product manager who owns Payment Guard's quality bar at Anchorwell. She'd written the original spec herself, back when the whole pitch to the founders was "a business owner should never feel our fraud check." For the first year, that pitch held up. Every Monday she pulled a sample of the fraud reports the model had missed and read them by hand, checking whether the miss was a shrug or a real problem.

She kept that habit for eleven months. Then the miss rate stayed steady at 46 percent, month after month, and every sample she pulled looked like the last one: small, forgettable, absorbed. So the Monday read became a monthly one, and by spring she mostly trusted the dashboard's summary line instead.

Then came Riverbank Landscaping. Fourteen employees, a client of Anchorwell for three years. Their equipment supplier's email got hijacked, and a message came through that looked exactly like the real thing: "Our bank account has changed, please update it before the next invoice." Riverbank's bookkeeper updated it. The next payment, $61,400 for a season's worth of mowers and parts, went out through Payment Guard.

Payment Guard didn't hold it. The amount wasn't unusual for that account. The payee was new, but new payees weren't enough on their own to trip the old bar, because the bar had been written to keep holds rare above almost everything else.

Same feature, two very different mistakes to make

Riverbank's owner called Anchorwell's support line the same afternoon the supplier called asking where their payment was. Abena opened a dispute case. Under UCC Article 4A, the bank only has to reimburse if its security procedures weren't commercially reasonable, and Payment Guard had, on paper, done exactly what it was built to do: hold almost nothing, bother almost nobody. Twelve weeks later, Riverbank got back $2,300. The rest, gone.

Abena pulled the year's full numbers that week for the first time since the habit had faded. Forty fraud attempts had slipped through in twelve months. The average business got back 9 cents on the dollar, same as Riverbank. Nobody at Anchorwell had been tracking that number as a single line anywhere, because the document that governed the feature only ever talked about how rare a hold should be.

Here's what I'd take back. The original spec said "hold under half a percent of payments," a ceiling, and it never once said what percentage of real fraud that ceiling was allowed to let through. Rewrite it, and the ceiling becomes a floor: catch at least 85 percent of confirmed fraud, and let the hold rate land wherever that actually requires. Run the same numbers with that bar and Payment Guard holds about 11 payments in every 1,000 instead of 3, and catches 89 out of every 100 fraud attempts instead of 54.

And the thing I'd tell myself, if I could go back to the spec review where "keep friction low" got written into the first line: a promise about how the feature feels is not the same thing as a promise about what it catches. We wrote down the one we could see. The one we couldn't see is the one that cost Riverbank their mower money.

PICK, spelled out for a quality bar

This is a tradeoff wearing a documentation section's clothes, so PICK is the tool, not a checklist of "things a good PRD includes."

The same pick, on a recycling truck

Cobblestone Waste and Recycling runs single-stream trucks for a mid-size city. A camera watches each hopper load for contamination, a trash bag, a propane tank, food waste, before the load gets compacted with the rest of the truck's route. Same question, a very different desk.

Here the position doesn't flip. The camera should still be a floor on what it catches, not a ceiling on how often it stops the truck, because a missed contaminant doesn't cost one stop. It costs the whole truckload, once it's already mixed in with six other neighborhoods' recycling.

P. Catch nearly every real contaminant, and accept that the truck stops more often for a false alarm than anyone would like.
I. A false stop costs the driver about four minutes to check by hand and wave the load through. A missed one costs the whole truck: nothing gets sorted out once it's compacted and delivered to the recycling facility.
C. The stop is cheap and visible, over before the next block. The miss is silent until the truck unloads hours later, and by then an entire day's collection from several streets is either landfilled or fined as one contaminated load.
K. This flips only if false stops start making trucks miss their pickup windows often enough to push routes into paid overtime, night after night. That's not happening yet, so today, the camera stays aggressive.

Flashcards (click a card to flip it)

Check yourself Score: 0 / 0