Something happens the moment someone shares their screen to walk through a presentation. Their face disappears. The other participants turn into an audience. The content plays out in a straight line instead of being worked through together.

From that point on the meeting is over as a conversation. What is left is a recording with witnesses.

Reading and interacting are two different activities, and they work badly at the same time. A video call is built for the second one: a small group, live, thinking out loud. If what you have is material to be read, send it up front as a memo or a short video. Then the meeting itself stays reserved for the part that genuinely needs everyone present.

Calling a meeting is a responsibility. You earn people’s attention by arriving prepared and by having the right people in the room. Slides on the screen are often how that preparation gets skipped: the deck carries the meeting so nobody else has to.

The test is simple. If your presentation survives being sent as a document, send it as a document.

Related

• Audio conversations promote a more equal distribution of speaking time and alignment of speaking patterns, which leads to better group performance than video meetings, despite the absence of visual cues

When you see a text a second or third time you recognise it, and that feeling of recognition is easily mistaken for actual understanding. This is the fluency illusion: the ease of processing something gets read as a measure of how well you know it. Rereading and highlighting raise how easy something feels right now (retrieval strength), but they barely build the deeper storage that remains once the text is no longer in front of you (storage strength). Take the source away, and the illusion of mastery collapses.

Passive techniques like highlighting and rereading feel productive precisely because they ask for almost no cognitive effort. You register that something matters, but that is not the same as integrating it into what you already know: explaining why it matters, saying it again in your own words, or finding it later without the text. Without that step you stay at the level of collecting information instead of building knowledge you can apply yourself.

The fix is not a better way of highlighting, but building in friction: forcing yourself to retrieve, explain or generate something without the source at hand. That feels harder and less successful in the moment, precisely because you can fail at it, and that friction is what leaves a lasting trace. The same illusion is not limited to studying: nodding along through a text or a codebase because everything feels familiar does not give you the ability to solve the problem yourself either.

Related

• Rephrasing a text helps me to better understand what the author means

You’re on holiday and something breaks at work. You drive back, you fix it, everyone applauds. And you feel it too: you were needed, and it worked.

That applause is where it goes wrong. Praise is a signal: it tells the whole organisation what actually gets rewarded. You learn that you’re the one who always shows up. Your team learns that the call is yours, not theirs. Your own boss learns that you’ll handle it.

This isn’t ego, or hardly ever. It’s identity. If you’ve been the reliable one for years, stepping back doesn’t feel modest. It feels irresponsible. Which is why deciding to delegate more isn’t enough on its own: you can tell your team to take more ownership and still sit on the context they’d need to act on it.

So prove your commitment some other way. Share context before there’s a crisis. Say out loud who gets to decide what. And next time, before you get in the car, ask your colleagues one question: what do you need to handle this without me?

Your body has a natural chronotype: a morning type, an evening type, or something in between, driven by the circadian rhythms of cortisol and testosterone. It determines when in the day you perform best, physically and mentally, and the gap between a clear morning type and a clear evening type runs up to eight hours. Evening types reach their aerobic peak around 20:00, roughly eleven hours after waking up, while morning types get there around noon.

Training at the wrong moment costs measurable performance. A morning type who trains in the afternoon can perform 5 to 6 percent below their natural peak. For an evening type forced to train very early, the risk is a different one: the body has not yet reached its optimal temperature and neuromuscular activation, which raises both the injury risk and the perceived effort, especially on too little sleep.

Even so, consistency outweighs timing. Pushing a clear evening type into early morning workouts can backfire, but training at a non ideal hour still beats not moving at all. Your chronotype is largely genetic and does not shift radically. An owl does not turn into a lark by getting up early for a few weeks. Light exposure and a consistent daily rhythm can nudge it, but only gently.

You finish the day with thirty things still on your list. You ask your AI assistant to prepare a summary of the week’s meetings to send to the team. You give it five minutes, trust the output, and hit send.

The next day, three replies. None positive.

What went wrong? Not the AI. The AI did exactly what you asked. The problem is what you asked, how you asked it, and what you expected without saying so.

The human buffer

When you delegate to a person, they fill in the gaps. They read the context, ask questions, assume what you probably meant. They patch your incomplete instructions with their experience and goodwill. Decades of working with people have trained you to rely on this buffer. You learned to delegate just well enough that a capable human could save you from yourself.

So you never had to confront the actual quality of your instructions. The person you delegated to quietly filled in what was missing, and the work got done. You assumed you were a decent delegator. You probably were not. Neither am I.

AI strips away the buffer

AI assistants do not do this. They execute. They take your instructions at face value and produce output that reflects, without any filter, exactly what you asked for.

That is genuinely useful information. If the output is wrong, the instruction was wrong. And now you can see it.

Before AI, bad delegation was invisible. The person you delegated to absorbed the cost of your vagueness. They spent extra time figuring out what you meant, made judgment calls you never knew about, and delivered something that looked like what you wanted. The sloppiness was hidden in the process.

With AI, the sloppiness becomes the output. And it lands in your inbox, or your client’s inbox, or your team’s inbox, before you catch it.

This is not a technology problem. It is a delegation problem that technology has finally made visible.

What good delegation actually requires

Delegating well to a person or to a machine requires the same three things.

First: you need to know what you want. Not in vague terms (“a good summary”) but specifically. What is the purpose of this summary? Who is reading it? What decision should it inform? What should it not include? If you cannot answer these questions before delegating, the person or the AI will have to guess. Humans are better at guessing. That is their advantage. It is also why you never noticed you were not answering these questions.

Second: you need to explain it clearly. Knowing what you want is not enough if you cannot translate that into a brief that someone else can act on. This is the craft part of delegation, and it is a skill most people have never deliberately practiced. We learn it accidentally, from working with patient colleagues who trained themselves to read our minds.

Third: you need to know how to verify the result. Before you send anything, you need a moment of “does this actually do what I said I needed?” That verification step requires you to reconnect with your original intent, and it only works if that intent was clear in the first place.

None of this is new. It is just more urgent now.

Why everyone suddenly needs this skill

Delegation and management were once skills for people who had teams. You needed direct reports, collaborators, employees before these competencies became relevant to your daily work. Most knowledge workers navigated their entire careers without seriously developing them.

That changed.

If you work with AI in any meaningful way, you are now managing something. You are setting direction, communicating intent, and evaluating output. The same skills apply, and the same gaps get exposed.

The democratization of AI is also the democratization of management. Every individual contributor who uses an AI assistant is, in some small way, now a manager. And most of them are starting where managers have always started: thinking they are better at it than they are.

The good news is that the feedback loop is faster. When you delegate to a person and the output is wrong, it might take days to surface. When you delegate to AI and the output is wrong, you see it in seconds. That speed is a gift, if you use it as a learning signal rather than as evidence that AI does not work.

Making it better

The path forward is not complicated, but it requires being honest about where the problem actually lives.

When AI produces bad output, resist the instinct to blame the model. Ask instead: what did I actually ask for? Read your prompt as if someone else wrote it. Is it specific? Does it include the context the assistant would need to make the right judgment calls? Does it describe what success looks like?

Then revise the prompt, not as a workaround, but as a genuine attempt to articulate what you want. This is the work. It is uncomfortable because it forces you to think more carefully than you are used to before you delegate. But it is also exactly the habit that will make you better at delegating to people, not just to AI.

Over time, you will get faster at it. You will develop a sense for what information needs to be in a prompt versus what can be assumed. You will learn where AI needs explicit context and where it can be trusted to fill in reasonable defaults. This is not prompt engineering in the technical sense. It is delegation skill.

Want to go deeper?

If you want to work through this more systematically, I teach a delegation matrix in my KENSO masterclass: a practical tool for deciding what you do yourself, what you delegate to a person, and what you hand to AI. The masterclass is in Spanish, but the framework applies regardless of language or workflow. You can find it at kenso.es/masterclass.

Living in Catalonia as a Dutchman means I navigate four languages daily: Dutch, Catalan, Spanish, and English. When it comes to my devices, though, I have a clear preference: Dutch. That’s the language in which I think most deeply, and I want my interface in my native tongue.

That preference ran into a wall when we installed a HomePod. Since Siri doesn’t support Catalan, my wife and I settled on Spanish as the household language for it. What I hadn’t anticipated was what that meant for every other device in the house: all members of a HomePod household must use the same Siri language. My devices had to follow the HomePod’s lead and use Siri in Spanish, even though my OS remained in Dutch.

Then Apple Intelligence arrived and made things more complicated: system language and Siri language now had to match. I tried a Spanish OS for a while, but eventually concluded that Dutch simply works better for me and decided to skip Apple Intelligence rather than compromise.

During a recent clean reinstall, I set up my devices fresh without restoring from a backup. A few days in, I noticed something unexpected: Siri was responding in Dutch. It had defaulted to my system language during setup, and apparently nothing was blocking it anymore. The HomePod restriction I had struggled with for so long seemed to have quietly disappeared, possibly a change that came in with Apple Intelligence, even if I am not using it.

A small discovery with a surprisingly big impact on daily comfort.

Niklas Luhmann, the prolific sociologist behind the Zettelkasten method, often described his note-taking practice as having a conversation with his system. He would write a note, and the system would “respond” by surfacing related ideas, unexpected connections, unexpected tensions.

It was a metaphor. A beautiful one, but still a metaphor.

But what if it didn’t have to be?

At the PKM Summit in Utrecht last week, I proposed a breakout session around exactly this question: what would it mean to have literal voice conversations with your personal knowledge management system? Not just speaking to transcribe, but speaking to think, and having your notes respond.

I’ll be honest: I went in with more questions than answers. And I came out with even more. But the discussion surfaced a framework that I find genuinely useful.

The four levels of voice interaction

The participants in our session quickly recognized that “voice with PKM” means very different things depending on how sophisticated the interaction is. We mapped it into four levels.

Level 1: transcription

This is where most people start. Tools like Wispr Flow, Onit, or Handy let you speak and get text back. It’s a fast, friction-free way to capture thoughts, especially useful when your hands are busy or you think faster than you type.

Useful, certainly. But this is essentially just a faster keyboard. Your PKM isn’t doing anything with your ideas yet.

Level 2: AI post-processing

Here’s where it gets interesting. You speak your thoughts, including instructions, and AI restructures them into something more useful than raw transcription.

This is what I do myself. I’ll verbally draft an idea in Tana, weaving in instructions like “turn this into bullet points” or “reorganize this around the main argument.” The result isn’t just captured speech; it’s been processed into a form my PKM can actually work with. Non-linear thinking becomes structured notes.

The key difference from level 1: you’re not just inputting; you’re collaborating.

Level 3: interactive voice dialogue

At this level, the AI doesn’t just process your input, it engages. It asks clarifying questions. It pushes back. It says: “You mentioned three different things here; which one is the core point?”

This is where voice starts to feel like the conversations Luhmann was describing, and where I’m currently experimenting myself. Using the voice chat feature in Tana, I’ve found that my thoughts become sharper not because I transcribed them, but because something challenged them. It’s slower than level 2, but noticeably deeper.

Level 4: the Jarvis mode

This is the aspirational level. Imagine being able to ask your PKM: “What did I decide about this project in January?” or “What sources have I read on habit formation?” and getting a coherent spoken response drawn from your own notes.

Full integration. Your PKM as a genuine thinking partner you can speak with, not just write to.

We’re not there yet. But the direction is clear.

The tension I can’t resolve

Here’s what I keep coming back to.

A few years ago, I wrote a post arguing that writing is the best medium for deep thinking. The act of writing forces you to formulate ideas clearly, to make vague thoughts concrete, to notice where your logic breaks down. The friction is the feature.

Voice removes that friction.

So is voice input actually good for thinking, or just good for capturing? Is it a shortcut that helps you move faster, or a shortcut that lets you skip the hard part?

I don’t have a clean answer. My instinct is that levels 1 and 2 are primarily about capture efficiency: getting ideas into the system faster and with less friction. Levels 3 and 4 are where real thinking might happen, but only if the AI engagement is substantive enough to replicate what writing does naturally.

The question is whether today’s tools get there.

Where are you?

Most people working with PKM systems are operating at level 1. A smaller group has reached level 2. Levels 3 and 4 are still being figured out.

What I took away from the PKM Summit is that the ceiling is much higher than the current tools suggest. And that the interesting work isn’t just building better transcription. It’s figuring out how spoken dialogue can become a genuine mode of thinking, not just a faster mode of typing.

Luhmann’s metaphor might become literal sooner than we think.

At which level are you using voice in your PKM? And do you think voice can ever match writing as a medium for deep thought?

Multitasking is a common phenomenon in our modern society, but research shows that we are much worse at it than we think. The problem lies not only in the fact that we struggle to perform multiple tasks simultaneously, but especially in our inaccurate assessment of our own capabilities. Studies show that people who frequently multitask wrongly believe they are doing so effectively. They significantly overestimate their abilities, while the reality is different. This lack of self-awareness is worrying: we are not only bad at multitasking, but we also seem unable to see or acknowledge it.

The consequences of frequent multitasking go beyond just a distorted self-perception. Research on chronic media multitaskers, people who regularly process multiple media streams simultaneously, shows striking results. These heavy multitaskers are actually more sensitive to environmental distractions and irrelevant information in their memory. Paradoxically, they even perform worse on tests measuring the ability to switch between tasks, likely because they are less capable of filtering out distracting information. This suggests that frequent multitasking is associated with a fundamentally different way of processing information.

The comparison to a drunk person who thinks they are walking in a straight line captures the problem well: we think we are better than we actually are. Thus, multitasking creates a dangerous combination of reduced performance and inflated self-confidence. As multitasking plays an increasingly larger role in our daily lives, it becomes all the more important to be aware of these limitations and the gap between our perception and reality.

Related

• The human mind and brain lack the architecture to perform multiple tasks simultaneously
• Attempting to do two or more attention-demanding tasks simultaneously reduces productivity
• Whether you think you can or think you can’t, you’re right
• Avoid excessive task switching to increase efficiency