Presence You Can Ignore: Engineering Non-Intrusion in a Shipping Desktop Companion
WUKO is a small character that lives on your Mac’s desktop and sometimes speaks first. That can easily turn into nagging. This paper shows how we built it so you can ignore it: it has a home spot, moves slowly, its unprompted remarks fade on their own, and nothing keeps score if you don’t answer. It is just as plain about the three places where the design falls short.
Prefer the original? Read the full paper · doi.org/10.5281/zenodo.22806802 →
The paper, in plain words11 min read
Why a desk companion is hard to get right
A chat window is a place you go to. You open it, you ask, it answers. Every chat starts with you, however clever the assistant is. It can be helpful. It cannot be company.
WUKO is a small character that lives on a Mac desktop and keeps you company while you work. To be company, it has to be able to speak first. But then it can interrupt. So how can it start a conversation without becoming a nuisance?
The app ships in two versions, one from the App Store and one downloaded directly from the maker.
The usual answer is good manners. But software cannot hold a mood the way it can hold a number. So the paper builds the quiet out of numbers. They set where the character may be, how fast it may move, and what its remarks may ask of you. Interruptions cost real things: slower work, more mistakes and more stress.
WUKO’s unprompted lines have no job. They are a joke, a riddle, the time, a passing thought, a memory or a kind word.
When a remark wants nothing from you, there is nothing for you to fail at by ignoring it.
Think of a friend who mentions the weather while getting on with their afternoon. You owe them nothing. Advice leaves a small debt if you ignore it. The design aims for the friend.
It has a place on your screen, not a window
A window is something you open. A character is already there. So WUKO is a borderless shape with a home spot in a corner.
At rest it floats above ordinary windows. But it does not take your clicks. They pass straight through everywhere except the character’s own body. Click its face and an arc of seven buttons opens right where you were looking.
How it moves: slowly, like breathing
A companion that stops moving looks like a crashed program. So the character bobs gently up and down, a little like breathing.
By default the bob makes one rise and fall every four seconds. That is 15 a minute, like fifteen breaths a minute, inside an adult’s resting range. Mood changes the speed. Across every mood written for it, the bob stays between about 9 and 25 a minute, as Figure 1 shows.
About 57 a minute would look agitated, so the top of the range was set well short of it. That reasoning survives only as a note in the code. Nothing there enforces it.
The range covers what is written, not everything the character can do. The AI behind it can call a tool that speeds the bob up 1.7 times, to about 43 a minute. The code’s only hard limit is 2 hertz (2 bobs a second, or 120 a minute).
Why slow motion is the right bet
The paper does not claim that slow motion escapes your attention. One study of working screens sorted motion into two kinds. Anchored motion stays around the object’s own spot. Travelling motion leaves it. People rated travelling much more distracting, and slow back-and-forth among the least irritating. A character bobbing in place is the first kind.
That study rated how distracting motion is, not how much it grabs attention. Its slowest test ran at 1 bob a second, so this band is two to seven times slower. The paper assumes the study’s result holds at these slower speeds. It did not measure them.
Other work finds that objects that loom (seem to rush toward you) or travel do grab attention. So the claim is narrow. Two of the three known kinds are absent: sudden appearances and looming. The third is the walk. Not every animation was checked for looming.
Picture a fish tank: you can rest your eyes in it. One catch. This calm is how the character acts around a stranger. As your relationship grows, it gets livelier.
Surprises should be rare
A perfectly regular presence fails too. Anything that happens the same way every time becomes an annoyance. So now and then the character has to do something. Even a surprise wears out if it comes the same way every time, and faster the more often it comes. That is why rarity is a setting, not a courtesy.
The paper found no study of where rarity stops feeling alive and starts feeling random. It states that as an open question.
Isn’t that a slot machine?
Rare, unpredictable events sound like a slot machine. The paper says the comparison is wrong, taken literally. A slot machine pays after a varying number of pulls, so the reward depends on what you do. WUKO’s big moments run on a timer, whatever you do. Nothing you do changes the odds, nothing is scored, and missing one costs nothing.
The paper admits that a glance and an event can coincide, and that the way this holds attention is shared with persuasive design. But the paper argues that compulsion needs unpredictable timing, a link to your behaviour, and something at stake. This design has only the first.
The paper turns this into a rule it calls I-VT. A noticeable event comes on a timer and never depends on what you do. No reward, score, streak or unlock, and nothing records whether you saw it.
The shipped odds are small. The biggest moment has a 1-in-250 chance per idle line. A chase invitation has 1 in 40. Whistling happens at most twice a day, with a further 1-in-3 roll. That unpredictability helps here is asserted, with a reason given. It was not measured.
When it goes for a walk
Now and then the character announces a walk and leaves its corner. That is travelling motion, the distracting kind, so it is fenced in, and the fence has gaps. One switch in the care panel turns the automatic walk off. It is on by default.
Here the fence holds. Over a true full-screen app, the rule is built into how the character is drawn. On a walk it shrinks to at most 60% of its size and drops to a lower layer. Anything below 85% of full size cannot sit in front of such an app.
Here it does not hold. Over ordinary windows, the character floats above them in 37.5% of walk stretches when all eight walk activities are on offer. On a bare desktop only four are on offer, and the share is 25%. The walk never checks which window you are using, so it can head straight for it. Only one of the eight activities steers around windows. Five ignore them.
One exception is deliberate, and the paper calls it the breach. A walk must be running, the character must be over a full-screen app, fifteen seconds must have passed since the last such reaction, and a one-in-three roll must come up. Then it rises, says one line, and fades back in 7 to 9 seconds. The paper calls it an attention grab, and did not measure its cost.
A message you can ignore
When the character speaks unprompted, you can walk away and nothing happens. The line stays for 12 seconds by default, then leaves. Nothing in the app can cancel that timer.
Two things follow. First, there is nothing to dismiss. Tapping a passive line does nothing, and the close button appears only when a reply box is open. No action is available. Second, nothing counts your ignoring. There is no ignore streak, no time spent looking, no read receipt and no response rate. The counters that exist count only what you did, such as riddles solved.
Questions are different on purpose. An ignored question stays open, and the next thing you type within 180 seconds counts as the answer. If you have hidden the character, a remark arrives as a Mac notification, which must be cleared.
What happens when you say no
A system that speaks first has to decide what your silence means. The paper’s starting point, asserted rather than proved, is that one ignored bubble is unclear. You may have been mid-sentence, or seen it and let it pass. Backing off hard on silence is guessing. A declined bubble is clear. So only a refusal changes the wait.
This looks like a bug, but it is not. In two places an ignored ask comes back sooner than a declined one. The permission ask returns in two days after silence and in seven after a refusal. Asking again after silence is a second chance for someone busy.
Saying no is a ladder, drawn in Figure 6. Ignore it, and nothing happens. Decline it, and the wait grows, then grows again. Switch the channel off, and it stops. Or quiet everything with one gesture.
One shared timer holds the rule: one day, three days, a week, then silence for good after four refusals. The permission ask keeps its own schedule, so the rule holds across much of the system, not all of it.
Controls switch things off rather than turn them down a little. A switch makes a promise you can check. A knob set to 0.3 could not be checked without watching for a long time. Each channel also has a dial, for example for how often idle lines come.
The project’s checker watches 78 pools of lines, with 7,627 written lines and no duplicates. The true total is larger, since the checker misses some files. Two gaps remain. Seven channels have no switch of their own, and nothing in the app tells you that declining buys more quiet than ignoring.
What it watches, and what it says
The paper counts five channels where the character speaks about what it sees. The walk names open apps and your desktop files. A break monitor times your work from how you type and scroll. A habit digest uses which apps you switch between. Two smaller ones notice app switches and first launches.
The line is drawn at the mouth. It names apps and, on the walk, your desktop files. It also remarks on your wallpaper and counts your screenshots. It never speaks a window title or the text of a document. Its keyboard monitor records only when you type, never which keys, and nothing leaves the machine.
Some lines still sound like the character sees more than it does. It can say “you’ve reread that line three times” when the code knows only which app is in front.
It goes quiet when you leave, not when you work
There is no typing check and there are no quiet hours. The reading window before background lines stop was lengthened to eight minutes, so a person who is reading is not mistaken for absent. One exception: on a call or while streaming video, it goes silent on its own. It recognises eleven calling apps and seventeen streaming patterns.
So it goes quiet where the signal is reliable, and does not guess where it is not. A named call app is clear. “Deep in writing code” has no such signal, and a guess would make the companion feel dead or rude. In the paper’s words, it did not protect focused work. It protected against performing to an empty room.
Figure 7 shows the break monitor’s timing: an eye nudge at 20 minutes, a movement break at the interval you pick, and a hard stop at two hours.
What the design does not refuse
It would be handy to say the app has no engagement metrics, the scores apps use to bring you back. Only part of that is true. Refusing costs nothing, and no score ever goes down. But the relationship side does use them.
It has a fourteen-level score. It has a daily-login streak with a bonus of up to four times that resets if you miss a day, and a care streak. A collection counter includes unearned secret items as placeholders. The paper places the app among companions that take a relationship role, at its gentle end.
It does not punish absence. The relationship score never drops. Close the app for a month and it opens at the level you left. But that holds only because no code lowers the score. A change could break it without a test failing.
One part is not defended: a needs loop. The character’s health levels fall in real time, from full to empty in roughly a day, even while the app is closed. The paper says its thesis does not rest on this.
Two things cannot be refused. The first-run tour cannot be skipped. And in the App Store version the character wears a graduation cap until you buy, and grumbles about it at most once a day. You cannot switch off a hat and a joke about a hat. You can switch off, in four taps, being asked for money.
The paper names the unflattering reading: that the grumbling uses the character’s distress to sell. It argues the tone is comic. It has not asked any users.
What the paper does not claim
The paper lists its own limits. These matter most:
- One app, one user (the author), and no controlled evaluation.
- The walk is the distracting kind of motion, and the breach’s cost was not measured.
- There is no limit on the total attention it asks for. Rarity is set per channel, not for the whole system.
- Whether the trial grumble reads as comic or pushy is a question about users, and none were asked.
So the paper does not claim that its motion never grabs attention, that its surprises feel alive, that the app is free of engagement metrics, or that it protects focused work.
What it proposes, and does not test, is that presence which never needs a reply may feel different from presence that does, and that this may matter to someone working alone. Testing that needs a group of people and someone other than the author. If you were designing the test, which of the three places where the design does not hold would you measure first?
The DOI is the citable artifact and Zenodo holds the PDF. This is the concept DOI, so it always resolves to the newest deposit.
The design this paper describes ships in WUKO, the lab's local-first desktop companion for the Mac.
WUKO →