Design a feedback system for a product I name.

The interviewer names the product: VetLoop, an ambient AI scribe. It listens during an exam and drafts a clinical note, findings, assessment, plan, for the vet to review, edit, and sign. Dr. Marisol Feng has practiced small-animal medicine for six years at Kestrel Hollow, sharing an exam-room tablet with two colleagues.

How to answer this, stage by stage

Nobody is grading whether you can name a feedback button. They're grading whether your design still works on the day someone would rather stay quiet.

Let's learn

The tablet in Kestrel Hollow's second exam room gets passed between three vets a day, still warm from whoever used it last.

Before VetLoop, writing up a full exam note took Dr. Feng about twelve minutes per patient, typed between appointments. VetLoop cuts that to under two: it listens during the exam and drafts the note, findings, assessment, plan, for her to check and sign.

Four steps, and the third one is where every bit of useful information currently disappears.

Right now, with no feedback design built at all, a wrong draft just gets fixed silently, the same way you'd fix a typo, and the correction goes nowhere else.

The turn: the silence was never about vets not noticing errors. Dr. Feng noticed every one of them; she's good at her job. The silence was about what reporting one would have said about her.

Only one of these two paths depends on anyone feeling brave. The other one happens either way.

At its worst: the same rare-condition note error, VetLoop confidently writing an incorrect assessment for an uncommon presentation, gets independently caught and quietly fixed by several vets across the network, none of them telling anyone, for months, until it finally reaches a patient record uncorrected somewhere it shouldn't have.

What I would leave alone: a genuinely one-off typo or phrasing quirk doesn't need any of this. The automatic diff will show it as noise, and it should stay noise; the design exists for the patterns, not for every single keystroke.

The lesson: a feedback system that only works when people feel safe enough to use it isn't really a feedback system. It's a hope that nobody will ever have a reason to stay quiet.

Now here is the same thing as a story

The short version above is what you'd pitch to VetLoop's product team. Read this one for how the gap actually got found.

Dr. Marisol Feng has practiced small-animal medicine for six years, the kind of vet who can catch a subtle limp before an owner even mentions it.

For VetLoop's first several months, she used to mention it casually, in the break room, when a draft came out strange: "VetLoop wrote something weird about the Hendricks cat again." It was a joke, mostly, and it doubled as an accidental feedback channel nobody had designed on purpose.

VetLoop kept confidently misjudging the same rare, uncommon-in-general-practice condition, and as the joke wore thin, the "AI errors" log rolled out clinic-wide, reviewed monthly, sorted by name. Reporting an error stopped feeling like a joke and started feeling like an admission that she needed the tool's help too much.

Five months of quiet fixing, and the story from a colleague is the only reason anyone connected the pattern.

At a regional veterinary conference, Dr. Feng ran into a colleague from a partner clinic in the same network, who mentioned, almost in passing, that a VetLoop-drafted note on a similar rare case had made it into a permanent record uncorrected, and had come up during an insurance audit.

Dr. Feng realized, standing there, that she'd caught and silently fixed the exact same kind of error at least five times over the past few months. If she'd said something the first time, an engineer might have caught the pattern before it ever reached her colleague's patient record at all.

Run the same six months forward under the redesign: every one of those forty-one edits gets diffed and logged automatically, no vet ever needing to decide whether reporting makes them look bad. The pattern crosses a detection threshold by the ninth occurrence, network-wide, weeks before Dr. Feng's colleague's case ever reaches an audit.

I put the quality log out in the open, sorted by name, because transparency sounded like the responsible choice at launch. It took one conference conversation to see that a design depending on people feeling safe enough to admit a mistake isn't transparent. It's just quiet in a different, more dangerous way.

SPARK, in one screenNot a features list. SPARK is what forces a feedback design to survive the day someone would rather say nothing at all.

Four fields on every edit, and the fourth one is what makes the whole design survive silence.

The recap, one line per letter: situation is silent fixing with no trace, payoff is never having to decide whether to admit a mistake, anchor is automatic diffing plus an unattributed optional tap, risk is the design surviving total silence from any one vet, and keep out is the heavier tools deliberately left for later.

And if you want to be sure it really works, try it somewhere elseSame five letters, an HVAC dispatch tool instead of a vet clinic. Nothing else about the two jobs is alike.

Briarfield HVAC Services uses an AI tool that drafts a diagnosis and repair estimate for a technician to review before a customer sees it. Marguerite Bissette has dispatched technicians there for five years.

Mapped onto SPARK: situation is the same shape, a technician quietly correcting a wrong diagnosis before sending it to a customer, with the correction going nowhere else. Payoff: never needing to decide whether flagging a wrong estimate makes them look like they can't do the job without the tool. Anchor: the same two-part design, an automatic diff of every changed line item, plus one optional tap for "this mattered," unattributed in any shared view. Risk: a technician embarrassed by a recurring miss on a specific unit type stays silent forever, and the diff still surfaces the pattern regardless. Keep out: no mandatory written report, no named leaderboard, at least not yet.

The two that made the cut both sit in the same corner: almost no effort, and still genuinely useful.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "capture every edit automatically, add one unattributed tap, never make anyone decide whether admitting a mistake is safe," and stop.
Cost: engineering says automatic diffing across every note costs a real build sprint. Say so, and start with the optional tap alone for v1, since even a partial signal beats none, while the automatic diff gets built next.
The model gets better, for real: if VetLoop's overall accuracy improves, this design still matters, because a rarer miss is even more likely to feel like an individual vet's fault rather than the model's, which makes silent concealment more likely, not less.

Three real ideas, and every one of them would have cost more trust than it earned on day one.

Where people run it wrong.
They build a feedback channel that only works if someone volunteers embarrassing information about their own reliance on the tool.
They make the "responsible," transparent choice of full attribution, without noticing what that attribution actually costs socially.
They wait for a support ticket or a formal complaint, which a quietly concealed error, by definition, never generates.

How to use it live. When someone asks you to design a feedback system, ask yourself first: does this still work on the day the person using it would rather nobody knew? If the answer is no, the design isn't finished yet.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0