Design the settings page that gives users control over an AI feature.

Seraphina Kolb has subscribed to Wavelight for three years. Most nights, once her kids are down, she watches for about an hour, and autoplay has always been the thing that let her not touch the remote at all.

How to answer this, stage by stage

You're not being asked to list every possible setting. You're being asked which one decision the whole page should hang on.

Let's learn

Picture the one hour after the kids are asleep, and everything Wavelight assumes about how you spend it, correctly, for months, until one night it doesn't.

Wavelight is a streaming service. Its AI watches what an account plays and skips, then queues up the next thing and recommends more of it, without anyone asking.

For most of three years, Seraphina finished an episode and the next one simply started, no remote required. She built a real habit around that: she stopped even holding the remote most nights.

The turn: the bad recommendations weren't really the problem. The problem was she had no way to tell Wavelight "that wasn't me," so the only lever she had left was to stop trusting autoplay at all and search by hand every single night.

What I would leave alone: a single-person account with no shared logins doesn't need any of this. Autoplay drifting toward someone else's taste can't happen if there's only ever been one someone.

The lesson: a settings page for an AI feature isn't really about exposing every dial the system has. It's about giving the person exactly one lever, in the exact moment the system gets it wrong, so they never have to quietly opt out of the whole thing instead.

This is the routine the whole design has to protect, without asking her to give up the part that made it effortless.

Now here is the same thing as a story

The short version above is what you'd say pitching this settings page in a design review. Read this one for the actual night that made the case.

Every night around nine, once the kids are down, Seraphina makes tea, sits in the one armchair that's hers, and opens Wavelight. It's the only hour of her day nobody else is asking anything of her.

For most of three years, that hour ran itself. She'd finish something, the next episode would already be starting, and she never once thought about the remote sitting untouched on the arm of the chair.

Nothing broke on one bad day. It drifted for weeks before it ever became a moment she noticed.

There was no single bad day that started this. Her kids had started using her profile during the day, on the shared living-room screen, since there was only ever one login on the account. Slowly, over weeks, Wavelight's sense of "what this account likes" started drifting, a little kids' content mixed into her evening queue, nothing dramatic enough to notice at first.

Then one night, three episodes into her show, autoplay quietly slid into a children's cartoon. She sat there for almost a full minute before it registered, tea going cold, half-watching something meant for a six-year-old at eleven at night.

Nothing about that night was catastrophic. It was just the moment she stopped trusting the queue to know who she was.

It wasn't the cartoon that bothered her. It was realizing, all at once, that Wavelight didn't know her at all. It knew whoever had last touched the remote, and had no way of telling the difference, and neither did she have any way of telling it.

What that cost, at its worst: she started searching manually almost every night after that, giving up the one piece of the product that had made her evening actually feel easy, because the alternative, trusting a queue that might veer anywhere, felt worse.

Four small things on one screen, and every one of them exists because of that specific 11pm minute.

I shipped a single shared profile with autoplay always on, because it made the product feel effortlessly smart the moment anyone signed up. It took watching someone quietly give up the entire feature, rather than risk one more surprise, to see that effortless isn't the same thing as trustworthy.

All three would help someone, eventually. None of them would have caught the 11pm cartoon faster than a plain reason and an undo button.

With the reason tag and the undo button live, the same drifting pattern now surfaces after two or three kids' titles, not three weeks of them, and one tap on "undo this signal" resets her evening queue before it ever reaches an actual show.

SPARK, in one screenNot a feature list. SPARK is what forces the whole page to hang on one decision instead of ten small ones.

The recap, one line per letter: situation is her either accepting drift or giving up on autoplay, payoff is actively steering the taste profile, anchor is the three-control settings page, risk is a shared login quietly reshaping the queue, and keep out is everything that wouldn't have caught tonight's problem anyway.

And if you want to be sure it really works, try it somewhere elseSame five letters, a fitness wearable instead of a streaming queue. A different industry, and this time the anchor is a dial on effort, not taste.

Larkspur Health makes a fitness wearable whose AI auto-adjusts a person's daily step goal and workout suggestions based on recent activity. Ingeborg Halvorsen has worn one for four months.

Mapped onto SPARK: situation is that today, Ingeborg either follows whatever goal the app sets or ignores the app entirely on days it's clearly wrong. Payoff is her actively telling the app how hard today should feel, instead of silently overriding it in her head. Anchor is a training-intensity dial she can nudge for the day, plus a reason tag ("goal raised because of last week's pace") and a one-tap "pause auto-adjust for today" button. Risk is a bad week, an injury or a bout of flu, quietly training the app to expect less from her permanently. Keep out is a full manual training-plan builder, which almost no casual user would ever open.

The dial and the pause button do almost all the trust-building work. The export feature barely moves it.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "mood dials, autoplay off by default, reason plus undo on every row," and stop.
Cost: there's no budget this quarter for per-row reasoning. Ship the autoplay default change alone first, since it costs the least and removes the exact risk that caused the 11pm moment.
The model gets better, for real: even if Wavelight's recommendations get more accurate next quarter, a shared login can still drift the profile in a way accuracy alone never fixes.

Where people run it wrong.
They build a settings page as a list of every possible toggle, instead of the one decision the actual failure needs.
They ship a powerful rules engine nobody uses instead of the small, high-traffic controls that would have caught the real problem.
They leave a risky default, like autoplay on for a shared login, untouched because "most people are fine with it," which is true right up until they aren't.

How to use it live. When asked to design a settings page for an AI feature, ask yourself first: what's the one moment this feature will visibly get it wrong, and what single control lets someone fix that moment without losing the whole feature's convenience?

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0