The keyboard didn’t have a submit or enter button. When he handed me the phone, fuming, I noticed the green submit button was actually sitting behind the see-through keyboard. There was no button to hide the keyboard either. The message box filled the whole screen, so tapping anywhere just moved the cursor around his message, and every missed tap on the keyboard added another typo to the message he’d so angrily written.

After about a minute of this, we managed to scroll using the sliver of screen on either side of the message box, tapped some white space, and the keyboard finally went down. And voila, there was the submit button.

That was a minute of annoyance. But when something doesn’t work the way it’s supposed to, it can cost a lot more than a minute, which brings me to today’s topic: the panic bar.

You’ve likely seen a panic bar on your office building’s stairwell doors, at government buildings, museums, basically any large public space. You probably didn’t give it a second thought. How complicated could it be? How could it even have a history? It’s just a door that you push open. What more could there be to this? Well, unfortunately, a couple of hundred dead children.

A free show in Sunderland

It was a beautiful, bright and warm day on Saturday, 16 June 1883, in Sunderland, a port town in the north east of England. Alexander Fay, a traveling entertainer, was putting on a children’s show that afternoon at the Victoria Hall, a big brick hall that could seat around 2,500 people. Tickets were a penny, and about 2,000 children showed up, with very few adults to watch over them.

The hall had three levels: the ground floor, a middle level called the dress circle, and the gallery at the top. Around 1,100 kids sat up in the gallery. To get down, they had to take a winding staircase with a door near the bottom.

At the end of the show, the entertainers started handing out toys and books from the stage, to the kids on the ground floor. The gallery kids saw that and ran for the stairs, worried there would be nothing left for them.

But unfortunately for them, the door near the bottom of the stairs had been bolted into the floor, partly open. The gap was 1 foot 10 inches, just enough for one child at a time. So now you have 1,100 kids running down the stairs, all needing to get through a door that only fits one of them.

The kids at the front couldn’t stop, because everyone behind them kept pushing, and the kids at the back had no idea anything was wrong. One boy, William Codling, who was six or seven at the time, wrote about it years later:

Suddenly I felt that I was treading upon someone lying on the stairs and I cried in horror to those behind “Keep back, keep back! There’s someone down.” It was no use, I passed slowly over and onwards with the mass

Bodies piled up in the gap between the door and the wall, and the pressure was so strong that the bolt holding the door actually bent. For context, crowds have been known to push with over 1,000 pounds of force, enough to bend steel railings. When everything settled, 183 children had died, many of them just 7 or 8 years old.

A relief fund raised £5,769 (about $28,000 USD at the time) for the families, and Queen Victoria herself gave £50 (about $240). Side note - £50 from the Queen sounded low to me, so I looked it up: £50 back then was worth about £3,300 in today’s money (around $4,400), or around five months of a skilled worker’s wages. Still feels a bit low coming from the Queen, if you ask me.

The inquest (an official hearing a coroner holds to work out how someone died, when a death is sudden or unnatural) recommended that “all doors, both internal or external, open outwards.” Today there’s a memorial in a park across from where the hall stood: a marble statue of a mother holding her dead child.

Subscribe now

The panic bolt

Ok, so how do you make sure a door can always be opened from the inside? Today that sounds like a no-brainer, but back then a door was locked with a key or a bolt, and somebody had to be there to undo it. And theatres and halls had a good reason to keep their doors locked: people sneaking in without paying. At a theatre fire in Brooklyn in 1876, the side doors were normally kept locked to stop people getting in for free, and when the crowd got to one, it was locked with no usher in sight. Even the door in Sunderland was probably bolted half open on purpose, so the kids would come through one at a time and their ticket numbers could be checked.

So venues had a real problem: how do you keep people from getting in, while making sure everyone can always get out?

In 1892, a man called Robert Alexander Briggs got a British patent for “Bolts and Fastenings for Doors of Theatres and other Public Buildings.” You push a rod on the inside of the door, and it pulls back the bolts at the top and bottom, so the door swings open. He called it a panic bolt, and it was first fitted at a new theatre in Edinburgh later that year. One push opens it from the inside, while from the outside the door stays “absolutely secure from unauthorised entry.” Perfect. Jolly good, problem solved!

The family story goes that Briggs was so shocked by what happened in Sunderland that he decided it should never happen again, and eight years later he applied for the patent on the bolt he’d perfected.

A “fireproof” theatre in Chicago

Unfortunately, the lesson didn’t make it across the pond.

It was Wednesday, 30 December 1903, in Chicago. The weather was fair and freezing, with temperatures in the low teens Fahrenheit (around minus 10°C). The Iroquois Theatre, which had opened just five weeks earlier and been advertised as “absolutely fireproof,” was putting on a holiday afternoon show, and the audience was mostly women and children.

During the show, a stage light set the curtains on fire. And it turned out that fire safety just wasn’t much of a priority at the Iroquois:

• Only one door at the main entrance was open. The others were locked, “according, it appears, to custom.”

• Other doors were “either frozen or rusted” shut.

• Some exits were hidden behind drapes.

• Many doors had “bascule” locks, a European style of lock that most Americans had never seen and didn’t know how to open.

So people did what the kids in Sunderland did. They rushed the few doors that would open, and the crowd behind kept pushing. The Chicago Tribune wrote that “scores were wedged in the doorways, pinned by the force of those behind them” (dozens of people stuck in the doorways, held in place by the crowd pressing into them from behind). Bodies were found “six and seven feet deep.” Nearly 600 people died, most of them within about 15 minutes.

So, same as in Sunderland: in an emergency, people couldn’t get through the doors they were pushing against.

The salesman who skipped the show

Carl Prinzler was a manager at a hardware store in Indianapolis. The story goes that he was actually in Chicago on a sales trip, planned to catch a show at the Iroquois, and changed his plans at the last minute to head home.

When he heard what happened, he was angry at “the hardware men who stood by idly with no solution.” He’d noticed that people rushing for an exit “usually fall at the doors themselves, many piling up against the doors.” In his own words: “By that time we felt something MUST be done!”

So he teamed up with his neighbor, an architectural engineer, and they came up with a bar that runs across the door. The first one went on a door at an Indianapolis high school in 1908, five years after the Iroquois fire, and the patent came through the year after.

They started selling it in 1910. A year later, a fire broke out at another Indianapolis school, and every student got out within 45 seconds.

What I find interesting is that both of these stories come back to regular people who saw a tragedy and felt like they had to do something about it themselves: a young man in England and a hardware store manager in Indianapolis. It reminds me of this quote:

“Everything around you that you call life was made up by people that were no smarter than you.” - Steve Jobs

Old habits die hard

You’d think that would be the end of it, but it took a lot more disasters for the rules to catch on:

• 1911, New York: 146 garment workers died in the Triangle Shirtwaist Factory fire. The law already said factory doors should open outward “where practicable,” and the ninth floor stair door, the one workers couldn’t get open, didn’t.

• 1913: America’s fire safety association set up a Committee on Safety to Life, which said exit doors should swing the way people are escaping. Its work became the Life Safety Code that US buildings still follow today.

• 1929, Paisley, Scotland: 71 children died at a New Year’s Eve cinema show after a film canister started smoking. There wasn’t even a fire. The doors opened inward, and the bodies piled up behind them.

• 1942, Boston: 490 people died in the Cocoanut Grove nightclub fire. One exit actually had a panic lock, but it was “rendered useless” by a second lock next to it that was locked.

• 2006: the US building code tightened the rule. Theatres, schools and other rooms for 50 or more people now need panic bars, down from 100.

• 2024: Dollar General paid a $12 million settlement in the US over stock blocking emergency exits in its stores.

So when did all of this actually become law? There’s no single date, because it happened city by city. London required theatre doors to open outwards as early as 1879, and in February 1892, the same year Briggs got his patent, it ruled that exit doors could only be fastened with “automatic bolts,” the push-to-open kind. Funnily enough, Chicago already had a rule that theatre doors open outward before the Iroquois fire. Less than three weeks after the fire, the city added that exits couldn’t be hidden behind drapes or locked while a theatre was open. And by 1939, America’s national exit code said schools and theatres needed “fire exit bolts” that open with no more than 15 pounds of push.

Today, in the US, any room that holds 50 or more people needs doors that swing outward, and in the UK the cutoff is 60 people. The bar has to cover at least half the door, and it still has to open with no more than 15 pounds of push, the same number as back in 1939. Remember the 1,000+ pounds a crowd can push with? The bar only needs 15.

A door that tells you what to do

So how does it actually work? When you push anywhere along the bar, it pulls back the latch holding the door shut. The door opens to “any horizontal force on the bar.” You don’t need to grip or twist anything, or know how it works. Your hands, your hip, or someone falling against it all do the job.

What makes it such good design? Don Norman, who wrote The Design of Everyday Things (I wrote about it in good design is hard to notice), has a set of design principles, and the panic bar ticks pretty much every box. The Iroquois theatre in Chicago, on the other hand, failed a lot of them:

• Discoverability (can you even figure out what actions are possible and where?): the bar runs across the door at waist height, so it’s hard to miss. At the Iroquois, some exits were hidden behind drapes. And at a nightclub fire in Rhode Island in 2003, three other exits were open and lit up, “but most people never realized they were there.” 100 people died.

• Affordance (what an object lets you do with it): Norman says “a flat plate mounted on a door affords pushing.” And pushing is exactly what a crowd does anyway.

• Signifier (the signal that tells you where to do the action): the bar is its own signal. You don’t need a sign telling you to push it. Norman says that when signs “have to be added to something as simple as a door, it indicates bad design.”

• Mapping (how the thing you touch relates to what happens): you push toward the outside, and the door swings toward the outside, the same way you’re trying to go.

• Feedback (letting you know your action worked, which I wrote about in how errors can lead to a positive experience): the door starts opening the moment you push, so you know straight away it’s working.

• Conceptual model (a simple explanation in your head of how something works): you basically don’t need one. The European standard for panic bars says they have to work “without prior knowledge of operation.” The bascule locks at the Iroquois needed exactly the knowledge nobody had.

• Constraints (limits on what you can do, so it’s hard to get wrong): the bar only moves one way, in. There’s nothing to twist, pull or figure out.

That said, it’s not perfect. The panic bar is great at one thing: opening when you push it. It has no plan B. If the latch gets stuck, or something is blocking the door from the other side, nothing tells you what to do next. Do you push harder? Pull and try again? Give up and look for another door? And you can’t see what’s on the other side until you’re already pushing against it. This isn’t just hypothetical either. Remember Dollar General and the stock blocking its exits? And at a nightclub fire in Brazil in 2013, security guards briefly blocked the only exit because they thought customers were running off without paying their bar tabs. The best door in the world can’t help if someone else decides it shouldn’t open.

A side note, and something ironic about the name: the people who study crowds say panic is mostly a myth. I’ve personally never been in a fire or a crowd crush, so I don’t have a first-hand account, but according to crowd science, panic is rare, and in emergencies people tend to help each other. Crowds become deadly because of pressure. When too many people are packed into too small a space, they physically can’t control where they move anymore, and people die from being squeezed until they can’t breathe, not from being trampled. Or as one crowd scientist put it: “People don’t die because they panic. They panic because they are dying.”

So the next time you push open one of those doors, maybe take a moment to think about the people who lost their lives simply because this wasn’t a thing when they needed it. I know I will.

But how?

First, think of an app or service where you ran into a problem. It doesn’t have to be digital, it could be in person too. Like the bus being late, or your food delivery saying “arriving in 10 minutes” for half an hour.

Recently, I opened an account with a new bank and was waiting on my debit card. The local shipping company messaged me to pick a delivery date and time (I needed to be home to receive it). The day came and went, and I heard nothing. I waited a day, thinking maybe they missed the window and would come the next day at the same time? Nothing. Waited another day. Still nothing.

On the third day I called them, and there was no option to speak to a person… It was all pre-recorded messages pointing you to their WhatsApp. And when I messaged them there, there was no live chat either. It said “we’re connecting you to a support staff, please share your details so we can assist you”, but after I typed out my message and all my details: nothing! It was like the chat froze. The negative emotions flooded in: frustration, annoyance, anger, and even hate!

(Side note: I ended up having to call the bank to get them to contact the shipping company and sort it all out! Terrible experience overall, and I didn’t even need the card urgently…)

But why did I feel so strongly? Because I didn’t feel in control. Imagine if, on the day of the delivery, they had messaged me this:

“Hi Ammaar, we’re sorry for the inconvenience, but we’ve had to reschedule your delivery to 14th October between 5pm and 7pm. Your package is now marked as a priority for the next delivery. You can also change your delivery slot or switch to pick-up by clicking here. Rest assured, timely delivery of your package is our top priority. Thank you for choosing [company name].”

I would have felt completely different. Sure, I might be a little annoyed if I’d cancelled my evening plans that day, but I would feel in control. I’d know the delivery was delayed and when I’d get it, and I could pick a new time or switch to pick-up. I wouldn’t be kept in the dark.

Earlier this year I was in Kelowna, Canada, and would take the bus to university. Like every other university student, I’d often catch the bus that got me to campus JUST in time. Sometimes strollers or wheelchairs getting on and off would delay the bus by a minute or two (which feels like forever when you’re waiting for a bus you needed to be on 20 minutes ago). But that never spoiled the bus for me. I never dreaded the ride. Thanks to the transit app, I could see the live location of the bus… I was in control. I wasn’t kept in the dark.

So a big part of how we experience a product comes down to feeling in control.

Why does feeling in control matter so much?

Don Norman (the cognitive scientist who wrote the book) describes exactly what happened with my debit card. When things don’t go as planned we get frustrated and angry, “especially when neither the reason nor the possible remedies are known.” I didn’t know why my card hadn’t shown up, and I had no way to fix it.

Then he says this: “Feedback provides reassurance, even when it indicates a negative result. A lack of feedback creates a feeling of lack of control, which can be unsettling.”

A message saying my delivery is late is still bad news, but it’s bad news I can plan around. The bus was late too, but I could watch it coming.

So how do we make people feel in control?

A few ways. First, some definitions from the book:

• Feedback: communicating the results of an action.

• Conceptual model: a highly simplified explanation of how something works.

• Mapping: the relationship between the elements of two sets of things (like which light switch controls which light).

Feeling in control needs good feedback, a good conceptual model and good mapping.

Good feedback tells you that your action was received, so you’re not left in the dark guessing whether it went through.

Think of a washing machine or a microwave. If you hit a button and no light came on and there was no sound, you wouldn’t know if the machine got it. You might press it again, which could actually turn it off if the machine did get it the first time! The machine received the input, it just didn’t have the right feedback to tell you. Cue frustration, annoyance and all the rest.

A good conceptual model helps you understand what’s going on and sets your expectations. In my shipping example, my conceptual model was that once I got the message saying my delivery was scheduled, the whole system was automated. So if the delivery got delayed, I’d get an automated message letting me know.

We usually pick up conceptual models from past experiences, and in this case it’s Amazon. Amazon may or may not have made shipping a lot harder for every other company, because you’re always getting updates on your order: when you place it, when it ships to the local center, when it’s out for delivery, and when it’s delivered. So now we expect that from everyone.

Domino’s did the same thing with pizza back in 2008, with a tracker that shows your order being made, going in the oven, and heading out for delivery.

And good mapping gives you better intuition and better expectations. In the transit app, I can see a map with the bus icon, the route, and where I am. It’s an intuitive mapping: the bus on the map is where the bus is in the real world. Since that mapping is accurate, I have a good idea of where the bus is and how long it’ll take to get to me.

Now think of a row of light switches in a corridor. Say the first switch is for the light closest to you, but the second switch is for the light furthest away, and the third is for the one in the middle. That’s a bad mapping, and you’ll get annoyed every time you flick the wrong one.

Back to the error

Errors will always happen. Buses run late and deliveries get rescheduled. Whether you walk away annoyed or happy mostly comes down to whether you felt in control while it happened. If that shipping company had sent me that one message, I’d probably have walked away impressed with them.

Has a company ever handled a mistake so well that you liked them more after?

As for the writing equipment: I use a cheap orange mechanical pencil I bought from Dollarama. I’ve used mechanical pencils ever since I was first introduced to them. To me, they’re quite the product. The 0.5 / 0.7mm lead (by lead, I mean the thin grey stick inside that you write with, not the metal, more on that later) means the tip is always sharp, which gets rid of the need for a sharpener, and the mess that comes with it.

Since it’s still a pencil, you can erase any mistakes you make while writing. The only downsides are running out of lead refills at an awkward time and the rubber on the back finishing quickly.

All this got me thinking about who invented the mechanical pencil and how it came about! Turns out every version of it was someone fixing one annoying thing about the version before. Here’s what I found.

Problem 1: you needed a knife

Before we actually talk about the problem, a quick detour: there’s no actual lead in a pencil, and there never was! In the 1560s, people in Borrowdale, a valley in the north of England, found a deposit of really pure graphite. It looked and marked like lead, so that’s what everyone called it. In 1779 a chemist showed it was actually a form of carbon, but by then the name had stuck.

Ok, so back to needing a knife. In 1565, a Swiss naturalist called Conrad Gesner described “a sort of lead... shaved to a point and inserted in a wooden handle.” He even drew it, and it already looks a lot like a mechanical pencil. But you still had to push the lead along and sharpen it yourself, with a knife.

Pencils needed a knife for a long time after that. Then in December 1822 (over 250 years later!), two Londoners, Sampson Mordan and John Isaac Hawkins, got a British patent for “improvements on pencil holders.” The patent says the whole point was to make “the frequent cutting or mending of the nibs or points unnecessary.” Their pencil worked like a glue stick: you twist the end of the pencil, and a threaded rod inside turns and pushes the lead out a little at a time. Twist pencils like this stayed the most common type for over a hundred years.

These weren’t cheap, though. Mordan’s were made of silver, and Victorian pencils were treated like jewelry, worn on a chain alongside scissors and sewing tools. The male guests at Queen Victoria’s wedding in 1840 were even given gold ones as gifts.

(Mordan wasn’t technically first, though. Divers found a piece of a sliding pencil in the wreck of HMS Pandora, a British ship that sank off Australia in 1791.)

Problem 2: it was fragile and fancy

Ok, now we need to take a trip to Tokyo, almost a hundred years later, where Tokuji Hayakawa ran a small metalworking shop. His first hit was a belt buckle that held the belt tight without needing any holes.

In 1915, a big pen company ordered a load of metal parts from him for their mechanical pencil. The pencil had a thick, clunky casing made of celluloid (an early type of plastic), and it broke easily. It was basically an expensive toy. Hayakawa was only hired to make the parts inside, but he thought he could make the whole pencil properly useful, the way a fountain pen was.

So he redesigned it. He replaced the mix of small parts inside with a single piece of brass, which made it much tougher, and swapped the plastic casing for nickel-plated metal. A year later, his sales agent gave it a catchier name: the Ever-Ready Sharp Pencil.

Shops in Japan didn’t want it at first. They said it didn’t go with a kimono, and the metal felt cold to hold in winter. (Most people in Japan still wore traditional clothes back then, and pencils had only just started being made there, so a shiny metal pencil felt out of place.) Then World War I left Europe short on supplies, a trading company in Yokohama placed a big order, and the pencils started selling overseas. “Sharp pencil” is still the word for a mechanical pencil in Japan today.

Then on September 1, 1923, the Great Kanto earthquake hit Tokyo. Hayakawa lost his two children and his wife in the fires. He was in debt, so he handed the pencil business and its patents over to pay it off, moved to Osaka and started again. In 1925 his new company made Japan’s first crystal radio, and he named it Sharp, after the pencil. That company is still around today as Sharp, a Japanese electronics company that sells TVs, microwaves and other electronics in well over 100 countries. You may have heard of them (I hadn’t!).

Around the same time in the US, Charles Keeran patented a pencil in 1915 that he called the Eversharp (the two names get mixed up a lot, but they were separate pencils). It was the first one made in huge numbers with a simple mechanism and room for lots of spare lead. By 1921 the factory was making 35,000 a day and had sold over 12 million. So the mechanical pencil went from a silver wedding gift to something a lot more people could own.

Problem 3: the lead was thick and snapped

The Eversharp’s lead was about 1.2mm thick, more than twice as thick as the 0.5mm in most pencils today, and that became the standard size. Going much thinner wasn’t really an option, because of what the lead was made of.

Since 1795, leads had been made using a method from a Frenchman called Nicolas-Jacques Conté. France was cut off from English graphite at the time, so he came up with a new way to make leads: grind the graphite into powder, mix it with wet clay, shape it into sticks and bake them. It worked, but clay is brittle, so a really thin lead would just keep snapping.

Jumping ahead 165 years, in 1960 the Japanese company Pentel launched what it calls the world’s first lead made with plastic resin instead of clay, in 0.5, 0.7 and 0.9mm. Swapping clay for resin is what made leads that thin strong enough to use. The same year it released a click pencil that, in Pentel’s words, never needs its lead sharpened. That’s basically the pencil I use today. Now 0.5 and 0.7mm are the most common sizes.

(The H and B on leads, by the way, are usually read as hardness and blackness. HB, right in the middle, is the standard one for writing.)

So how does the click actually work? Think of it like pulling a rope towards you hand over hand: grab, pull, let go, reach back, grab again. Inside the pencil there’s a tiny metal claw holding the lead.

• Press the button: the claw moves forward, carrying the lead with it, then opens and lets go.

• Let go of the button: the claw slides back up. A small rubber ring at the tip holds the lead still, so it doesn’t slide back too.

• The claw grabs the lead again, a little further up.

Each click repeats that, so the lead moves forward a tiny bit at a time. This short animation shows it really well.

Problem 4: thin lead wears down lopsided

If you’ve used a mechanical pencil for a while, you’ve probably noticed the tip wears down at a slant. Mitsubishi Pencil, the Japanese company that makes uni-ball pens (and no relation to the Mitsubishi that makes cars), says this makes your lines get thicker, makes the lead break more easily, and leaves crumbs that smudge the page.

Their fix was the Kuru Toga in 2008. It has tiny gears inside that turn the lead a little with every stroke, so the point wears evenly and stays sharp. It’s sold over 100 million. You can see the gears turning in their video (also, it’s refreshing to see a launch video from before AI that isn’t heavily edited).

Other companies went after the snapping instead. Zebra’s DelGuard has springs that absorb pressure so the lead doesn’t break, and Pentel’s Orenz has a metal tip that slides back as the lead wears down, so you can keep writing without clicking.

Here’s the timeline we just covered:

Back to my two problems

So after all that, my two problems from the start are still there.

Running out of lead hasn’t really been fixed. In 1978 the Japanese pen company Pilot made a pencil you shake to push the lead out instead of clicking, and in 2017 Pentel made one that feeds the lead out for you as you write. But with all of them, you still have to buy refills and load them yourself.

And the eraser has been an add-on from the start. Putting one on the end of a pencil was patented in the US in 1858 by a man called Hymen Lipman. He sold the patent for $100,000, and then in 1875 the Supreme Court threw it out, because it was just two things that already existed stuck together lol.

The pencil went from silver and gold to under a dollar at Dollarama in about 200 years.

(This would be the perfect spot for one of those photos, but I couldn’t find them!)

In 2023, while I was in Canada for university, my parents moved to Dubai. After graduating earlier this year, I moved back to spend some time with them. Apart from it being a new city to me, living here has kind of normalized the building. Depending on where I’m going, I see it far away in the skyline almost every time I leave the house, whether it’s to play padel or to go to the supermarket.

So I got curious about how tall it really is. Turns out a big chunk of it isn’t really usable building at all...

Vanity height

The Burj Khalifa is 828 meters tall. But the highest floor anyone actually uses is at 585 meters, The Lounge, an observation deck on floors 152 to 154. Everything above that is what’s called vanity height, the part of a skyscraper above the last floor people can use.

On the Burj, that’s 244 meters, or 29% of the whole building! It’s mostly the thin pointy tip on top, plus a few floors for maintenance, communications and equipment.

In 2021, Emirates airlines filmed an ad with a stuntwoman dressed as one of their flight attendants standing on the very top of it. It took the crew over an hour just to climb from the 160th floor to the tip at 828 meters. See the ad here.

So why build it?

Because it counts. The Council on Tall Buildings and Urban Habitat, the group that decides which building is officially the tallest, measures to the “architectural top”. That includes spires (the pointy tips) but not antennas, signs or flagpoles. So if you want your building to rank higher, a tall pointy tip is a pretty easy way to get there.

Before you jump on the “that’s so typical”, “it doesn’t count”, “that’s not fair” bandwagon, it’s actually pretty common! As of September 2023, the 100 tallest buildings in the world had an average vanity height of 12.6%.

One World Trade Center in New York is a famous example. Its roof is at 417 meters, and the mast on top takes it to 541. In 2013 the council ruled that the mast was a permanent part of the design, so it counted, and that’s what put it above Chicago’s Willis Tower (442 meters) as the tallest building in North America. The top 155 meters of it is vanity height.

(Also, to be fair to the Burj, even without its empty top it would still be the tallest building in the world, but only by about 2 meters.)

Putting 244 meters in perspective

What’s kinda crazy and blew my mind is that if you cut off just that empty top and stood it up on its own, it would be taller than the tallest building in about four out of five countries in the world. That includes France, Italy, Switzerland, Argentina and South Africa.

(And before anyone brings up the Eiffel Tower, it’s 330 meters but it doesn’t count as a building. The council only calls something a building if at least half of its height is space people can use, and the Eiffel Tower is mostly open metal frame with three levels you can visit.)

It’s also taller than every building in Antarctica. But the tallest one there is about 15 meters, so that’s not saying much.

Subscribe now

Contrary to your friends' remarks, you're not an idiot! Instead, you were a victim. A victim of bad design.

(Fun fact: a door like this is called a Norman door, named after Don Norman, a cognitive scientist whose book The Design of Everyday Things is one long complaint about stuff like this. He was also the first person to have "User Experience" in his job title, at Apple in 1993.)

So why is it hard to notice?

Because it just does its job without drawing any extra attention. To me, a watch with a strap is good design. I don't notice it and it's straightforward to use. Every morning I put on my watch and it takes less than a few seconds.

Meanwhile, something that does draw my attention is my WHOOP band! The standard band the device comes with (the one in the image below) took me about 5 minutes to figure out. The puck can actually be forced in place both ways, and the metal case can rotate a full 360 degrees. Even after a few months of having one, it still takes me 15 seconds of focused attention to figure out how it all goes together.

The cost of good design

You don’t need big instructions, simpler layouts or to give up on aesthetics to have good design. You can have good design and still keep the look and the feeling the product gives you top of mind. For example, with the glass doors, you can put a flat metal plate on the side you push and a handle only on the side you pull, and nobody needs a sign.

This has actually been in the news. Car companies spent years moving everything onto big touchscreens because they look clean and modern. But when it starts pouring and you're digging through menus on a screen just to find the wipers, with your eyes off the road, it’s not so ideal. So since January 2026, Euro NCAP (the group that crash tests and safety rates cars in Europe) only gives a car 5 stars if it has physical controls for the indicators and hazard lights, the wipers, the horn and the SOS button! I'm personally a big fan of this!

What are some examples of good and bad design from your life?

Leave a comment

But as AI got better, and out of habit too, I started using it more and more for tasks such as planning marketing experiments, pricing and deciding what to build next, slowly outsourcing my thinking. I mean pre-AI, I would just google something, watch a few YouTube videos and do what they advised me to, right? Isn’t this the same thing? So I started using it for strategy and planning, the more important stuff, and relying on the AI’s reasoning.

I think this is especially true for young professionals. We need AI to keep up with the pace of work, but we haven’t built up the on-the-job knowledge to actually use it as leverage.

AI in 2024 was at best a helpful assistant. It has since progressed and gotten closer to a point where it’s a solid sparring partner. But it hasn’t reached the point where it’s an expert and can guide you perfectly.

I think what’s important to understand is humans are quite complex. Most tasks are not quite as simple as getting from point A to point B like a GPS route. You can call this a “context” issue. AI can’t fully grasp our goals, emotions and circumstances, so it can’t guide us well.

Let me give an example. In my previous startup we were publishing digital worksheets and the goal was simple: increase usage. Pretty straightforward goal, right? I thought so too. This was early 2026 and Claude Code was getting seriously impressive. I was juggling multiple roles (i.e. busy like everyone), and thought if I created a specific folder (think project) in Claude with all the relevant context, some guiding principles and the goal to increase usage, Claude would handle it, right? I wouldn’t need to be too involved and could stay away from the battlefield, and just call the shots from afar: usage goes up, the startup takes off, we get rich.

Well, what happened was every week, I would give Claude all the results (i.e. context) from that week’s changes and tell it to analyze them and guide next steps. Every week Claude was PASSIONATELY CONVINCED that we were making progress, learning something new and that this week’s change would be IT. I was in the back seat, convinced Claude knew what it was doing, and let it do what it was doing for a few weeks.

At one point I recall it recommending a change with the same intensity as the previous ones, but it was so ridiculous that it made me spit my metaphorical coffee out. I pushed back, but Claude had been brainwashed, of sorts, and wouldn’t back off. It kept insisting it was a good idea.

This was the heart-sinking moment when I realized I’d just wasted the last four weeks. I thought I was being ultra productive but it was just slop. Well, not AI slop in its common form, but a more polished version: intellectual slop. Slop might be too harsh of a word, as it was genuinely trying to hit the usage goal. But it didn’t have the bigger layer of context: knowing when to call it quits, when the results weren’t meaningful enough, or when the goal itself needed a fundamental change. And that’s the difference from googling, right? The person in the YouTube video has actually done the thing, and knows when to stop.

The same logic applies to improving landing pages, writing a “good” cold email or even building a fat-loss plan. If you don’t have the domain knowledge and just rely on AI for the thinking, you won’t get far. You will be running on a treadmill with the illusion of getting somewhere.

So what should you do instead?

In my experience, the quickest way to get over this is to just acquire that knowledge firsthand. The best way to do that is good content. For the business world, this was YC articles, videos from credible people and last but not least: books. Reading a relevant book from an expert on a topic, absorbing the information in a way that works for you (I take notes and apply spaced repetition), and THEN using AI, is an insane multiplier.

You will be SO much more productive and efficient in reaching your goal. Some of you might have rolled your eyes when I said acquiring the knowledge firsthand, given that we’re in the age of voice dictating to our AIs and clicking enter, but it’s actually not that time intensive. Once you get over the hump of closing your AI app and picking up a book, it only takes like 5-6 hours to read most books. That’s 30 minutes a day for 10 days or so. Don’t forget to take notes or do whatever you need to do to actually remember the information you read. The return on your time spent reading will go materially down if you don’t absorb that info.

Once you have that knowledge, you will be able to jump past any bad ideas that the AI throws at you, have a better idea of what’s worth doing in the first place and have a granular understanding of the roadblocks you hit. You will also feel much better, as one does when they take control of their life lol.

Good luck :)

Subscribe now

How are you using AI?

Comment below and let me know. Alternatively, come chat with me.

I’ve got a bit of a history with cold emails. In the past few years, I’ve used cold emails to:

• 2023: land my first ever internship (yes, it was paid)

• 2024: get senior colleagues to mentor me at my second internship

• 2024: land another internship (in product management)

• 2025: partner with 9 apps in a particular niche (for my prev. startup)

  • Got a few CEOs to respond, which was cool

• 2026: get information on a niche program from incubators & VCs in a different country

If you’ve written cold emails in the pre-AI era, you’d know that it’s actually pretty time intensive to write these emails. Writing is also only one aspect of it: you also need to find relevant folks who are worth emailing in the first place, and research them. It’s like writing custom cover letters when applying for jobs. But of course, if it works, you’ve won the lottery.

For the first 4 of the bullet points I listed, I did most of the process manually. Googling, opening webpages in new tabs, endless scrolling and switching tabs. It took a LOOOT of time. But for that 5th bullet point, I got the same quality of results and invested a fraction of the time. That was because Claude Code + Chrome access actually got really good. So the manual research and custom emails that used to take 3 hours now happen in an hour, and most of that hour is writing a good prompt and reviewing the emails.

So, here’s how it works.

The results

In January 2026, Claude found 25 to 30 people for my list, each one with a source. I emailed about 20 of them, reading and sending every one myself. 7 replied, and several of those replies came with intros.

This one didn’t answer my question directly. It came the day after my follow-up, which I sent 6 days after the first email, and it pointed me to their founder community and the person who runs it.

She read my newsletter, subscribed, then named two organizations worth contacting.

Three organizations to contact in one reply, and each one became a new row on my list.

Why it works

Every email establishes credibility first, then asks for something small.

• The ask is small. A pointer or a referral. It’s specific and easy for them to say yes or no.

• The numbers are concrete. Real metrics, written once in a file so they never drift between emails.

• One line is personal. Every email has one genuine sentence about the person or company.

1. One folder, one brain file

Everything for the campaign sits in one folder. At the top is a file called CLAUDE.md, which Claude Code reads every time I open the folder, so I never re-explain anything.

campaign-folder/
  CLAUDE.md                   the brain
  Target List.md
  Weekly Progress Tracker.txt
  [Name] - Call Notes.txt
  Supporting Docs/
  .claude/settings.local.json

CLAUDE.md holds who I am, my numbers, the one approved email, the rules, and who I’ve contacted.

2. Claude builds the list

I give Claude 2 or 3 seed pages to start from, like an accelerator’s mentor page or a conference speaker list. It follows the links from there and reads LinkedIn profiles, team pages and articles in my own Chrome, where I’m already logged in. Simpler web tools get blocked on these pages.

It comes back with 25 to 30 people, one row each: name, role, organization, 2 to 3 sentences on why they’re relevant (with specifics read off the actual page), the best way to contact them, and the source URL.

The source URL matters most. Every personalized sentence in an email comes from that row, so Claude can’t make one up. If it can’t find anything specific about someone, the row says “no specific found.”

Different seed pages surface different people, so I run two or three passes from different seeds and remove duplicates.

3. One email, cloned forever

The approved email is four short paragraphs, under 200 words:

1. Who I am and my numbers, plus what makes them impressive (how fast, or what I was juggling alongside it).

2. What I’m doing and why them. This is where the one personalised line goes.

3. A small ask: a pointer or a referral.

4. Sign off.

I wrote the first one by hand and sent it for real. Only then did it become the template Claude copies.

4. Claude drafts, I send

Claude opens a pre-filled email in my mail app. I read it, tweak it, and send it from my own account.

Sending it myself means every email gets a human read, and every email lands in my real Sent folder. That second part is what lets Claude track replies.

5. Replies feed the machine

Claude checks my Gmail and logs every reply. After each one it:

• moves the person to “Responded”

• logs what they offered

• adds every name they mention as a new target

• starts the email to each of them with “[Name] suggested I reach out”

• saves anything the reply taught me as a new rule

A no or a hedge becomes a rule too. If someone declines a broad intro, the next emails ask for something specific, like “do you know anyone who has been through X?”

3 to 6 days of silence gets one follow-up, three sentences long. Then I leave them alone.

What actually got replies

Contact methods, best first:

1. A direct email, found publicly

2. hello@ or team@ addresses (these worked more often than I expected)

3. Web contact forms (low reply rate)

4. LinkedIn or X DMs (these don’t show up in Gmail, so I log them by hand)

5. Public posts in relevant groups

Start with the people who know everyone: connectors and program managers. They gave me the best replies, and their intros warm up the big names you email later.

The subject line says what you are, like “ADHD newsletter founder”. Never “quick question” or “intro”. A good hack I read somewhere is to word the subject line as if it’s an internal email.

The files

Six files do the work:

• CLAUDE.md: the brain. Profile, numbers, email, rules and log. Claude updates it after every send and reply.

• Target List.md: one row per person, each with a source URL. Claude fills it while browsing.

• Weekly Progress Tracker.txt: a four-week plan and weekly check-in. I update it when Claude asks.

• [Name] - Call Notes.txt: notes from any call the emails lead to. We both add to it.

• Supporting Docs/: my one-pager, CV, and anything I link or attach.

• .claude/settings.local.json: pre-approved Gmail and web search access. I set it once.

There’s no spreadsheet. Who’s replied, who’s waiting and who’s next all live in CLAUDE.md, so Claude checks it before drafting the next email.

Set it up yourself

You need Claude Code signed in with a Claude account (Pro, Max, Team or Enterprise) and the Claude in Chrome extension. If you add Gmail as a connector on claude.ai, it shows up in Claude Code automatically and Claude can check your replies. Without it, you paste replies into the chat.

1. Make a folder for the campaign and drop Cold-Outreach-Playbook.pdf into it.

2. Open Claude Code in the folder with claude --chrome and paste this:

Read Cold-Outreach-Playbook.pdf and set up this campaign with me.

3. Write the first email yourself and send it once for real. Only then does it become the template.

Cold Outreach Playbook

35.9KB ∙ PDF file

Download

Download

Subscribe now

How are you using AI?

Comment below and let me know. Alternatively, come chat with me.

In my experience, doing “ok” in courses is actually hard. I’m either doing very well and headed for a 90%+ or I don’t know anything and might just fail the course. This reached its boiling point in my last year of university when I was balancing working on my own startup with a full course load. After bombing a quiz, I went online searching for solutions and signed up for the iCanStudy program by Justin Sung. Although I didn’t get too far, I picked up just enough to do well in my courses with minimal time inputted!

I signed up for the iCanStudy program in September 2025, and started using Claude Code (the same engine behind Cowork) around the same time. Naturally, I applied the techniques I picked up within Claude Code and it worked very well. Techniques like spaced repetition and taking in content in different forms (practice questions, diagrams, simpler explanations) just so happen to overlap with what Claude Code is good at.

Alright, enough yapping. Here are my grades from my last year (2 semesters) of university:

For context, I worked 10 hours 3x a week during this period and used a version of the system that I explain below for COSC 328, COSC 320, and COSC 414 (the other two are capstone and a project-based course).

So how does it work?

1. The folder is the memory

Every course gets its own folder: lecture slides, syllabus, every past quiz and practice exam, assignments, and any emails from the prof. At the top of the folder sits a file called CLAUDE.md, which Claude Code reads at the start of every session.

CLAUDE.md holds everything a tutor would need to know: the exam format, what the prof said is on it, the topics I keep getting wrong, the topics I’ve locked in, and how I like to be taught. Nothing lives in my head or in a chat I’ll lose. I can skip a week, open the folder, and Claude picks up where we left off.

COSC 414/
  CLAUDE.md            the brain
  StudyPlan_Final.md
  Progress.md
  CourseContent_L1-3.md
  ReviewSheet_Final.txt
  Lectures/
  Example Questions/
  Quizzes/
  Archive/

2. Teach from the prof’s slides

This one has the clearest proof behind it. For my COSC 414 midterm, I had Claude teach me from summaries of the lectures, and I got 49%. After that I switched to having it teach from the actual slides, and I got 100% on Quiz 4.

Now Claude reads the real lecture PDFs a few slides at a time, in order, while I follow along on the same slides on a second monitor. One idea per message, then a quick check before moving on. Pages the prof said are excluded get skipped, and so do code slides if the course only tests concepts. Everything else gets taught.

It also uses the prof’s own method and notation, checked against the prof’s example questions. In 414 the prof tested the setup and never the arithmetic, so I practised writing out the matrices without multiplying them. Anything Claude pulls from outside the slides has to be labelled as such. Early on it taught me a lighting model from the textbook (Blinn-Phong) that was never on the slides, and that cost me a whole session.

3. Every session opens with a quiz

Before anything new, Claude quizzes me for 10 to 15 minutes on whatever is due, no notes. Every topic in the course has a mastery score from 0 to 5 in a file called Progress.md, and the score sets when the topic comes back:

• 0, New: review now

• 1, Introduced: in 1 day

• 2, Familiar: in 3 days

• 3, Confident: in 7 days

• 4, Mastered: in 14 days

• 5, Locked in: in 30 days

Get it right and the score goes up, so the topic stays away longer. Get it wrong and it comes back sooner. That’s spaced repetition, and it’s tedious to track by hand for every topic in three courses. Claude does the bookkeeping, so all I have to do is show up.

Questions come the way the exam would ask them, with no hints like “which pattern is this?” I write the answer and Claude checks it. Saying “yeah I get it” doesn’t get me out of the reps.

4. Practice like the real test

This one depends on your course. If past exams exist, they’re the most valuable thing in the folder. Claude catalogues every question in them one by one: what it tested and where it showed up again.

In COSC 320 (algorithms), the prof took “inspiration” from past final questions, and 2 of the 4 questions on the second midterm came from them. In COSC 414, every practice question showed up on the midterm, often one level harder.

Then I practise timed, the way I’ll write the exam. Claude writes a quiz into a text file, I type my answers in, and it reviews them. Every miss goes into a weak-spots table and decides what the next session opens with.

5. The system updates itself

Every session ends with the same checklist:

• Mark the session done

• Log what clicked and what didn’t

• Move the mastery scores

• Add to the cheat sheet

• Update the weak-spot table

• Save any corrections as rules

Then Claude lists every file it touched, so I can see nothing got skipped.

The rules are the part that compounds. Any time Claude did something that didn’t work for me, I told it, and it saved the fix. Over one term that added up to about 40 rules, like “one concept per message” and “screenshots, never ASCII”. That second one came from a midterm question I lost after studying from an ASCII diagram Claude drew in place of the actual slide.

The files

Six files do the work, and Claude keeps all of them updated:

• CLAUDE.md: the brain, read at the start of every session.

• StudyPlan.md: the schedule and a log of every session.

• Progress.md: mastery scores and review dates.

• CourseContent.md: the course, in the prof’s words, filled in as lectures are taught.

• ReviewSheet.txt: formulas, traps and mnemonics. Grows every session and becomes the cheat sheet.

• Quizzes/DayN.txt: Claude writes the questions, I type the answers.

Set it up yourself

The file attached to this post is the setup guide. It’s the instructions Claude follows to build all of this for a new course.

1. Fill a folder with everything from the course: slides, syllabus, past quizzes, assignments, prof emails.

2. Drop AI-Study-System-Setup.pdf into the folder.

3. Open Claude Code in that folder and paste this:

Read AI-Study-System-Setup.pdf and set up the study system for this course.

Claude asks you a few questions first (exam format and date, whether you get a cheat sheet, how many hours you realistically have), reads the whole folder, and builds the files. From then on you just gotta show up :)

AI Study System Setup

37.2KB ∙ PDF file

Download

Download

Subscribe now

How are you using AI?

Comment below and let me know. Alternatively, come chat with me.