What UX patterns make waiting feel productive rather than broken?
Latency budgets and UX tradeoffs
A silent five seconds on a voice assistant doesn't read as calm. It reads as a mic that stopped listening, and hands full of a knife go looking for something to say.
The direct answer
Waiting feels like work happening when four things are true: the person gets an instant sign they were heard, one honest marker lands partway through instead of silence, real progress shows up piece by piece instead of all at once at the end, and there's a way to check in without starting over. For a voice assistant, that's an echo chime under a third of a second, one spoken line timed to the exact second people usually start doubting it, a list that fills in item by item, and a mic that stays open so a nervous repeat doesn't collide with itself. None of these four patterns make the model faster. They fix the gap where a person decides whether to trust the wait or talk over it.
Do this, in order
Give an instant echo, a chime plus one paraphrased line, inside a third of a second of the request landing.Why: this is the single cheapest, highest leverage pattern on the list, and it's the one that answers "did it hear me" before the person even has time to ask.
Add exactly one honest spoken or written marker, timed to the real second people start doubting the wait, not a vague filler.Why: a marker that lands after people have already started repeating themselves is too late to stop the overlap it's meant to prevent.
Show real progress landing piece by piece, not dumped all at once when the whole thing finishes.Why: a growing list or a filling-in result is proof of work anyone glancing over can see, without needing to read a status line.
Keep the input open through the wait so a nervous check-in adds onto the same request instead of starting a second one.Why: this is the guardrail for when the first three patterns aren't quite enough. Without it, a repeat collides with the original and both get parsed as one garbled thing.
Say an uncertain guess out loud the moment it happens, not folded silently into a finished-looking result.Why: model output is a guess dressed as an answer until something confirms it, and hiding that guess inside a confident-looking result is how a small assumption turns into a bigger trust break later.
Skip all four patterns for anything that resolves in under about a second and a half.Why: wrapping ceremony around a near instant reply makes the fast thing feel slower than it actually is.
The menu this whole answer is built from, before any one product picks from it.
How to answer this, stage by stage
Nobody is grading whether you can name loading screen tricks. They're grading whether you know which of them actually earns a place in five real seconds, and why the rest can wait.
1
Scope it to one product and one real wait
Say it like this
"Let's ground this in Fetchpot, a voice assistant Cindergate built for the kitchen counter. Say 'find me a quick dinner from what's in my fridge, and add anything I'm missing to my list,' and it takes about five seconds to answer: check what fits, name a recipe, and build the missing items into your list. Ludovic Roark is the PM who owns those five seconds, and he cooks dinner with it most weeknights himself."
Why this works
A vague "design a loading screen" prompt turns into a list of tricks fast. One real five second wait, on a real counter, makes it a real design problem instead.
2
Open with the menu, not one trick
Say it like this
"This is a design question, so I'll use SPARK. Instead of jumping to one screen, I'll name the small set of patterns that genuinely make a wait feel productive, then say which of them earn a place in Fetchpot's five seconds specifically, and which ones I'm deliberately leaving out."
Why this works
Shows the interviewer you're picking from real options, not reciting the first idea that came to mind.
3
Reframe the question before answering it
Say it like this
"This isn't really asking me to name loading screen patterns. It's asking whether I know that silence, on a voice-only device, with someone's hands full, doesn't read as patience. It reads as 'did it hear me,' and the first thing a person does about that is talk over it."
Why this works
Separates a real judgment call from a features list nobody asked to be recited.
4
Give the anchor, the specific pattern combo and why
Say it like this
"Four things, in order. A chime and one paraphrased line inside a third of a second: 'Got it, quick dinner from the fridge.' One honest spoken marker at two and a half seconds, timed to the exact point our own numbers show people start talking over it, not a vague filler noise. The list fills in one item at a time as each gets confirmed, not all at once at the end. And the mic stays open through the whole five seconds, so a nervous 'are you there' adds onto the same request instead of colliding with it."
Why this works
This is the actual answer to the question. Everything else defends it.
5
Prove it with the failure, cut to four sentences
Say it like this
"Here's what happens without this. The first build had one steady pulsing ring for the full five seconds, covering listening and thinking as the same animation. On a normal Tuesday, Ludovic asked for dinner mid-chop, got no new signal three seconds in, and asked again. Fetchpot parsed both as one request and put a pasta dish and a paneer curry's ingredients on the same list."
Why this works
Shows the real cost of silence, not just the mechanism behind the fix.
6
Say what you'd watch after launch
Say it like this
"I'd watch the overlap rate, how often a second command lands before the first one finishes. Before this redesign it was thirty four percent of requests longer than three seconds. If that number doesn't drop once the patterns ship, the marker isn't landing at the right second, and I'd go recheck the timing before I'd add a fifth pattern."
Why this works
Shows you think about the design as something to check after launch, not something you ship once and forget.
7
Say what you'd leave alone
Say it like this
"Anything Fetchpot answers in under about a second and a half, like 'add butter,' skips all four patterns. A single quiet beep is honest there. Wrapping a spoken marker around a one second reply would make the fast thing feel slower than it is."
Why this works
Shows judgment instead of applying the same four patterns everywhere out of habit.
8
Close on the decision, not the story
Say it like this
"So: an instant echo, one honest spoken marker timed to when patience actually breaks, a list that fills in as it goes, and a mic that stays open so a repeat doesn't collide with itself. That's what makes five seconds feel like work happening instead of a mic that stopped listening."
Why this works
Ending on the reusable decision, not the anecdote, is what makes this sound like something you'd reuse, not a story you told once.
Let's learn
Every weeknight for years, Ludovic Roark spent about twelve minutes doing three things at once before dinner even started: checking what was actually in the fridge, scrolling a recipe app for something that fit, and writing whatever was missing on a scrap of paper.
Fetchpot is the voice assistant Cindergate built to do all three at once. Say what you're in the mood for, and it names a recipe that fits what's on hand, then builds whatever's missing straight into your shopping list.
Knowledge spark: what's an echo back?
A quick chime plus one short line repeating what the assistant heard, before it does anything else. "Got it, quick dinner from the fridge." It costs almost nothing to say and it answers the one question a person actually has in the first half second: did this thing hear me at all.
With Fetchpot, the whole thing takes under a minute: about five seconds for it to answer, then maybe twenty more to pick from the three recipes it names and glance at the list it built. Twelve minutes down to under one.
When people start talking over Fetchpot, by the second
Share of requests where the person has spoken again by that second
Only fourteen percent had spoken again by two and a half seconds, right where the marker fires. By three seconds, before any marker at all, that number nearly triples.
The extra second here or there inside that five second wait was never really the problem. Five seconds is a fair trade for twelve minutes, any night of the week. The real problem was that the five seconds looked exactly like nothing happening, and a person with a knife in one hand fills a gap like that with the only tool they've got: their own voice, said again, right on top of whatever Fetchpot was already doing.
A silent five seconds doesn't feel like patience. It feels like nothing happening, and a person with a knife in one hand fills that gap with their own voice.
Fetchpot's redesign, before and after, two error types
Before, single silent ringAfter, echo plus marker plus open mic
Both numbers moved together, which is what you'd expect: most wrong item lists trace straight back to an overlapping request.
At its worst, that overlap tangles two requests into one list, wrong items mixed in with right ones, and a second trip to the store to sort it out. Twenty five minutes lost undoing a mix up, on a task that used to take twelve minutes by hand and never tangled at all.
The choice I'd take back
We kept the same single, steady pulsing ring we'd shipped back when Fetchpot only answered one word adds like "add milk," finished in under a second. Nobody revisited it once five second recipe requests joined the one word adds eight months later, because the ring itself never changed, so nobody thought to ask if the job behind it had.
What I'd leave alone: anything Fetchpot answers in under about a second and a half, a plain add like "add butter," doesn't need any of this. A single quiet beep is honest there, and stacking a spoken marker onto a one second reply would make the fast thing feel slower than it actually is.
The lesson: a wait doesn't feel long because of the clock running underneath it. It feels long because the person waiting has nothing true to hold onto, so their voice fills the gap with its own guess about what's happening. Give the gap something honest to say, and the same five seconds stops feeling like a mic that stopped listening.
Now here is the same thing as a story
Read the story below when you want to feel why a light that keeps talking back earns more trust than a fast, silent one, not just be told that it does.
Ludovic Roark has cooked dinner for his family every weeknight for nine years. He knows what's actually in his fridge most nights before he even opens the door, just from habit.
Fetchpot showed up on the kitchen counter as a small speaker Cindergate sent every employee to test at home. The first two months were good. He'd say "find me something quick with what I've got," hear a chime, and in about five seconds Fetchpot would name three recipes and start filling in whatever was missing from the list. He'd glance at the light while he chopped an onion, half paying attention, mostly trusting it.
The habit thinned in three small steps. At first he watched the light the whole five seconds, curious. Then, after weeks of clean answers, he stopped watching and kept chopping, letting the wait happen behind him. Then one Tuesday, mid-chop, the wait felt a little longer than usual, and the light gave him nothing new to look at the whole time, so he said the request again, just in case.
There wasn't one bad night that started this. Nothing broke, not at first. It built up slowly, over about ten weeks, as Cindergate widened what Fetchpot could answer, one word adds giving way to full recipe requests that took real seconds instead of none. Nobody noticed the shift happening request by request. It showed up on a chart in the monthly product review instead: the share of requests where a second command landed before the first one finished had climbed from nine percent to thirty four percent, quietly, release over release.
Ludovic's own habit was inside that thirty four percent more nights than he liked to admit. On one of them, the light stayed on through his repeated request, and Fetchpot answered with a list that had a pasta dish and a paneer curry's ingredients tangled into one, six items he didn't need, at a store that only carried two of them. He caught it standing in the aisle, not at home, and drove back for what he'd actually meant to buy. Twenty five minutes gone, on a task that used to take him twelve.
We didn't lose five seconds to a bad guess. We lost a second trip to the grocery store, and a little of Ludovic's trust in a light he used to not think about at all.
It was never really about the five seconds. Nobody had a number for that. What they had was a light that had stopped telling the truth about what it was doing, and once it did, Ludovic's own voice quietly started covering for it, on more nights than the product review ever counted.
The decision worth taking back happened in a short design review, months before recipes were even on the roadmap. Someone asked what the ring should do while Fetchpot worked. Someone else said, reasonably, "keep it calm, one steady pulse, nothing busy," and that became the default, back when the only things Fetchpot answered were one word adds, finished in under a second. Nobody came back to revisit it when five second recipe requests joined the roadmap eight months later, because the ring itself never changed, so nobody thought to ask if the job around it had.
Run that same Tuesday the old way, and Ludovic's second request tangles the list again, same aisle, same drive back. Run it the new way: at zero point three seconds, a chime and "got it, quick dinner from the fridge." At two and a half seconds, before he'd normally start doubting it, "checking three recipes now." He keeps chopping. At five seconds, the list is already three items deep, filled in one at a time as Fetchpot confirmed each. No second request. No tangled list. No drive back.
One design gave Ludovic a light with nothing behind it, so he filled the gap with his own voice. The other gave him a light that kept talking back, so his voice stayed out of it.
What I'd tell the version of me in that design review: calm isn't a property of the ring by itself. It's a property of the ring matched to the job it's covering for. The same pulse that read as calm on a one second add read as broken on a five second recipe, and nobody checked which job it was doing anymore before they let it ride for eight more months.
SPARK, run for a wait you can only hear
Not a checklist to recite. Each letter has to survive the same overlapping request that the story just walked through.
SSituation. Who is this person, and how does the job get done today, without you?
Ludovic Roark, PM at Cindergate, cooks dinner for his family most weeknights and dogfoods Fetchpot doing it. Before Fetchpot, that meant checking the fridge by eye, scrolling a recipe app, and writing a paper list, twelve minutes across three tasks, one pair of hands.
Name the real person this wait has to hold the attention of, or the design floats free of anyone actually standing at a counter.
What Ludovic did alone, before any voice assistant touched dinner.
PPayoff. What habit do you want this to build?
Not "trust the silence because it usually turns out fine." Specifically: Ludovic hears the instant echo, keeps chopping through the marker at two and a half seconds, and only speaks again if Fetchpot actually asks him something.
A named habit produces a named set of patterns. A vague goal like "reduce anxiety" produces nothing anyone can actually design on a Tuesday.
AAnchor. The one design decision everything else hangs on.
Layer four patterns onto the five seconds, matched to a hands-busy, eyes-elsewhere moment. An echo chime and paraphrase at zero point three seconds. One honest spoken marker at two and a half seconds, timed to our own data on when people start talking over it, not a vague filler sound. The shopping list fills in one item at a time as each gets confirmed, real proof of work for anyone glancing at the counter. And the mic stays open the whole five seconds, so a nervous check-in adds onto the same request instead of colliding with it.
This is the actual design decision. If it doesn't visibly survive the next letter, it's a slogan, not an anchor.
Four patterns, each timed to a real second, not sprinkled in for decoration.
RRisk. What breaks the first time you're wrong?
Fetchpot's first guess is sometimes wrong, a swapped ingredient when something's out of stock, or a step runs a little long inside a busier request. If the design still stays silent past its usual second, an anxious repeat lands anyway, and two requests merge into one garbled list, the exact failure the story just walked through.
A design that only works when every recipe request finishes right on schedule isn't a design. It's a demo that hasn't met a slow Tuesday yet.
The exact chain the four patterns exist to break, right at the emphasized middle box.
KKeep out. What do you deliberately not build?
No fake animated percentage bar, recipe matching doesn't move at a steady rate, so a smooth number would lie. No five stage narration read aloud, that's too slow for a voice interface and would eat into the five seconds it's meant to be covering. Just the echo, the one marker, the growing list, and the open mic.
Piling on more spoken detail doesn't build more trust past a point. It just turns a five second wait into an eight second one.
Four patterns shipped. Two tempting extras didn't, on purpose.
Three things worth stating directly, since this is where the real judgment sits. The alternative rejected was a plain repeating chime with no words, tried first because it was cheap to ship. It lost because a sound carrying no information doesn't change when people start talking over it, the elapsed time curve looked the same with or without it. The AI specific failure worth naming by name is confident wrongness: Fetchpot sometimes swaps a missing ingredient for the closest thing already in the pantry, oat milk for almond milk, say, and if that swap gets folded silently into a finished looking list, it reads exactly as certain as everything else on it. The guardrail is saying the swap out loud the instant it happens, "swapped in oat milk, you're out of almond," never buried inside a list that looks equally sure about everything on it. And the trade off worth naming too: Fetchpot runs a small, fast model just to produce that first echo paraphrase, in parallel with the slower model doing the real recipe reasoning. That's a second inference call, a real cost, paid on every single request, to buy back the third of a second that keeps most people from ever needing to ask twice.
And if you want to be sure it really works, try it somewhere else
Same five letters, a photo of a dented fender instead of a spoken request, and this time the wait varies by how many damaged spots one picture holds.
Fenderscope is a feature inside Roadwell Insurance's claims app. Photograph the damage after a fender bender, and it reads the photo, checks it against your policy, and gives an estimate before a human adjuster ever opens the file. A typical claim takes about eight seconds. Malwina Absalon runs claims operations at Roadwell.
S, situation: before Fenderscope, every photo claim sat in a queue for a human adjuster, sometimes for days, and Malwina had never had to decide what shows on screen while an AI reads a photo, only reviewed an adjuster's notes once the whole thing was done.
P, payoff: the habit worth building isn't "trust the estimate because the app answered fast." It's a driver watching the damage regions get marked as they're found, and trusting the number because they saw how it was built.
A, anchor: the same four patterns, translated from voice to photo. An instant check mark the moment the photo uploads, "photo received, four angles good." One honest marker partway through, "checking your policy now," not a fake percentage. Damage regions highlighted on the photo one at a time as they're found, not one final overlay dropped on top at the end. And a low confidence region flagged the moment it's found, "this dent needs a closer look," instead of folded silently into a number that looks as certain as the rest.
R, risk: a scratch that reads as minor in one light and structural in another, and a confidently wrong estimate sent straight to a driver, who then argues with a number an adjuster has to correct later, worse than never showing a number at all.
K, keep out: no auto approval of payouts on day one, and no hand editing the marked regions. Just the marked photo and the honest estimate. A person still signs off before money moves.
The decision Malwina would take back
Early Fenderscope held every damage marker back until the very last frame, so a genuinely complex claim, six separate dents, looked frozen for the first six of its eight seconds with nothing new on screen. Drivers started retaking the exact same photo, thinking the upload had failed, before the first estimate had even finished building.
Same anchor, sorted by how long this particular photo actually needs.
Same method, a different weak spot: Fetchpot's wait varies by how many ingredients are missing. Fenderscope's wait varies by how many separate damage regions one photo holds, so the marker has to fire once early regardless of that count, not wait for the busiest photo in the batch to finish its first region.
Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the anchor, instant echo, one honest marker, progress landing piece by piece, an open channel for a check-in, and give the one number, overlap rate falling from thirty four percent to six.
Cost: there's no budget this quarter for a polished animated marker. Ship the plain spoken line first. The trust comes from the timing and the honesty, not the production value.
The model got better, for real: say Fetchpot's recipe matching gets faster and the whole wait drops to two seconds. That's not a reason to strip the echo back out. The echo answers "did it hear me," which matters at any length, it just means the marker itself may no longer be needed.
Where people run it wrong.
They animate a percentage that glides smoothly to one hundred regardless of what's actually happening, which teaches people the number means nothing.
They bury an uncertain guess inside a finished looking result instead of flagging it the moment it happens, so a small assumption reads as a confirmed fact.
They time the one honest marker to a round number that feels tidy, three seconds, instead of the actual second real people start doubting the wait.
How to use it live. Say the real tension out loud before answering: "is this asking me for a nicer loading animation, or for what a person actually does with their hands during the wait." That buys a beat, and it's almost always the second one.
Flashcards (tap any card to flip it)
1 · THE FRAMEWORK
What framework is this, and what's its one job?
Tap to flip
ANSWER
SPARK: design against the failure before you build. Here that means picking a menu of wait patterns and choosing the combo that survives the exact moment the first guess is wrong.
2 · THE PERSON
Who is this answer about?
Tap to flip
ANSWER
Ludovic Roark, a PM at Cindergate who cooks dinner for his family most weeknights and dogfoods Fetchpot doing it, nine years of knowing his own fridge by habit.
3 · THE HABIT
What did Ludovic stop doing once Fetchpot worked, then start doing again?
Tap to flip
ANSWER
He stopped watching the ring closely and let Fetchpot work behind him while he kept cooking, then started repeating requests once a wait felt even slightly off, with nothing on screen or in sound to say otherwise.
4 · THE ANCHOR
What's the one design decision the five second wait hangs on?
Tap to flip
ANSWER
Layer four patterns onto the wait: an instant echo chime, one honest spoken marker timed to the real second people start doubting it, a list that fills in item by item, and a mic that stays open through the whole request.
5 · THE OLD DECISION
What decision would you take back?
Tap to flip
ANSWER
Keeping the same single, steady pulsing ring from the one word add era, never revisited once five second recipe requests joined it eight months later.
6 · THE NUMBER
Fill in the blank: catching the tangled list cost Ludovic about ___ minutes standing in a store aisle, on a task that used to take ___ minutes by hand.
Tap to flip
ANSWER
Twenty five minutes lost, on a task that used to take twelve minutes total and never tangled at all.
7 · THE REPLAY
Same bad Tuesday, new design, what changes?
Tap to flip
ANSWER
At 0.3s, a chime and echo. At 2.5s, a spoken marker, before Ludovic would normally start doubting it. By 5s, the list is already three items deep. No second request, no tangled list, no drive back to the store.
8 · CROSS-PRODUCT TRANSFER
Section 4 answers this same question again for a different product. Which product, and what's the shared anchor?
Tap to flip
ANSWER
Fenderscope, a car damage estimator at Roadwell Insurance. Same anchor: instant echo, one honest marker, progress landing piece by piece, uncertainty flagged immediately, this time against a photo instead of a voice.
Check yourself Score: 0 / 0
True or false
1. True or false: the fix for Fetchpot's five second wait was making the wait itself shorter.
True
False
Show hint
Check the direct answer. It never says a word about making the model faster.
Show answer
False. Five seconds is a fair trade for twelve minutes by hand. The fix was an echo, a marker, a growing list, and an open mic, not a shorter clock.
Multiple choice
2. Which pattern is the anchor's highest priority piece, the one that has to land first and fast?
A. The spoken marker at 2.5 seconds.
B. The instant echo, inside a third of a second.
C. The shopping list filling in item by item.
D. Keeping the mic open the whole five seconds.
Show hint
Look at bullet 1 in the priority list and check what it says the echo answers.
Show answer
B. The instant echo answers "did it hear me" before the person even has time to ask, which is the cheapest, highest leverage fix on the list.
Fill in the blank
3. Before the redesign, ___ percent of requests longer than three seconds got a second, overlapping command. After, ___ percent.
Show hint
Look at the grouped bar chart in "Let's learn," the "Overlap rate" pair of bars.
Show answer
34 percent before, 6 percent after. The wrong item rate moved the same way, from 19 percent to 3 percent, because most tangled lists trace straight back to an overlapping request.
Short answer, name the rejected alternative
4. What alternative did the team consider for filling Fetchpot's five second wait, and why did it lose?
Show hint
Look at the paragraph right after the K step in the SPARK recap, where the three closing points are stated directly.
Show answer
Model answer: A plain repeating chime with no words, tried first because it was cheap to ship. It lost because a sound carrying no information doesn't change when people start talking over it, the elapsed time curve looked the same with or without it.
Short answer, apply it yourself
5. Pick an AI product you use yourself. Name one wait of a few seconds or more in it, and which of the four patterns you'd add first.
Show hint
Think of a feature where you've caught yourself tapping or repeating a request while it was still working.
Show answer
Model answer: A photo editing app's AI background remover shows one spinner for up to six seconds with no feedback. The first pattern to add would be the instant echo, a thumbnail outline of the subject appearing within a third of a second, so a person can tell it found the right shape before the full six second wait even starts.
Multiple choice
6. On the elapsed time chart, by what second had the share of people talking again already reached about fourteen percent, right where the spoken marker is timed to land?
A. 1 second.
B. 2.5 seconds.
C. 4 seconds.
D. 5 seconds.
Show hint
Check the dashed line and the labeled dot on the line chart in "Let's learn."
Show answer
B. Only 14 percent had spoken again by 2.5 seconds. By 3 seconds, with no marker at all, that number nearly triples to 38 percent, which is exactly why the marker is timed to land before that climb.
Before you close the answer
Why this works
Tests whether you notice this is an audio first, hands busy surface, not a screen someone is staring at. Most candidates design for a spinner on a monitor. Few design for a person whose eyes and hands are both somewhere else entirely.
Follow-up traps
"Doesn't the marker itself just add more talking, making the wait feel longer?" Response: one short marker timed to the exact second patience actually breaks is shorter than the alternative, a second full request getting reparsed and a garbled list needing a drive back to the store.
"What if the model genuinely takes twelve seconds sometimes, not five, doesn't the whole pattern break down?" Response: repeat the marker rather than invent a fifth pattern, and past a real length hand off to a person, the same move Fenderscope's K step makes for anything over eight seconds.
If pressed
The spoken marker at 2.5 seconds uses a slightly higher pitched tone than the echo chime at 0.3 seconds, so a person half listening while chopping can tell which stage they're in by ear alone, without ever looking up from the cutting board.
From U2xAI Academy
From answering questions to owning outcomes.
A live workshop where you ship a working AI agent, defend a launch decision, and walk away with a portfolio recruiters can't wave off, not just more questions to study.
