The Wrong Piano

The best-selling solo piano record ever made exists because a stagehand made a mistake. A teenager in Germany booked Keith Jarrett to play, and when he arrived, the piano was bunk.

Jarrett was twenty-nine. He had driven 350 miles from Zurich, hadn’t slept in two days, and was gritting his teeth in back pain. He had asked for a very specific piano for the concert, a Bösendorfer 290 Imperial. He got the baby grand the hall used for rehearsals. It was out of tune. Some of the black keys didn’t work right. The pedals were stuck, the top end was tinny, the bass was shallow. Only the middle register was really usable. It was not good.

Jarrett nearly bailed. But he looked at the teenager who had booked the event and the sold-out opera house and decided to go out and play anyway, knowing the performance was being recorded. Listen to the first ten seconds of the record. Jarrett plays four or five notes echoing the come-sit-down bell notes from the hall. The audience laughs, and then, at a broken piano in a dark hall in front of 1,400 people at 11:30 pm, with a back brace on and no plan, he played. The result was The Köln Concert.

He accepted the instrument in front of him anyway. Everything on that record had to be decided in the room.

Everyone in the hall knew it was an improvised concert. Jarrett had recently made that his signature product. In 1971 he recorded Facing You, a solo studio album that mixed pieces he’d written with sections he made up. Then he started playing solo concerts where he’d perform tunes and connect them with improvised passages in between. Then he dropped the tunes.

Jarrett had spent thousands of hours practicing songs, chords, rhythms, and harmonies. He ingested the design system of music. Then, armed with unique craft and taste, he branched off and created his own style of interacting with the 88 keys. The bravery to embrace possibility and “what if” solidified Jarrett’s brand.

When Jarrett decided to play the broken piano in Cologne on January 24th, 1975, the piano didn’t have instructions or warning labels about what was and wasn’t working. He just closed his eyes and listened, intuited, and responded. There’s a phrasing that starts around 4:50 and runs through the 9-minute mark that is at once so outside of any system and at the same time so perfect. Parts of the “design system” (chords, harmony, etc) are in effect, but the magic was Jarrett working within the constraints of that imperfection and producing something remarkable.

Sixty-six minutes, released exactly as played. No edits, no punch-ins, no “tomorrow’s show will be better.” Which means the whole thing is on the record, two records actually. Every place he reached for something that didn’t come, every stretch where he’s clearly circling and waiting for the next idea to arrive. He didn’t put a meeting on the books to get strategic clarity. He played through the unknown bars until they turned into the next thing, and some of what makes the record great is built on top of moments that lacked direction.

Those churning, hypnotic left-hand ostinatos that feel like weather rolling in are a workaround for a bass register that wouldn’t hold a note. You’re listening to a man solving a design problem in real time, in front of an audience. The broken piano is the shape of the music.

Your product is broken in places too. Some of those places you know about and you make work. Others aren’t as clear, and you don’t find out until you’ve hit a series of notes, reach for the closing one, and that key is stuck. Or the API doesn’t return information the way you planned for. Or the market wants something else entirely.

To augment some of these imperfections, we built design systems. Components, tokens, documented patterns, governance, adoption metrics. A drive towards perfection and control. The point was to stop redrawing the same button and free designers and engineers up to tackle harder problems. A design system is a tuned piano. You stop fighting the keys, and what comes out is what you already knew how to play.

A design system is like a book of tunes your team already knows, broken up into chorus, melody, and rhythm notations. For years we used these to stay consistent internally and go faster. Then AI models showed up that had read every song ever published. Now if you put a design system and AI together, they can produce a fluent, technically correct version of almost anything in an afternoon. The design, however, gets a little Truman Show, and everyone’s afternoon looks a little purple, a little Inter, a little “Here’s the part I keep coming back to.”

In contrast, Bill Evans and Keith Jarrett both recorded “My Funny Valentine.” Same song, same changes, two records that sound nothing alike. They both had 88 keys to work with. They shared everything a design system gives you: same harmony, same form, same tradition, both of them fluent in it. That more global design system didn’t keep them the same. What separated them was their personal pauses. Where they decided the tune could stand to lose something. Where they added a flourish. They each played their distinct version based on their own craft and taste, accumulated over thousands of hours of playing and listening.

Your product isn’t any different. Some people love Airtable, some love Coda, some build their own bespoke CRM. Each of them speaks to a different audience. Each has a different learning style or feature need, each contains a specific desire for control.

How many of us have heard:

Those requests arrived long before AI. They rarely produced anything close to the original. But copying used to be slow. Someone had to spend six weeks actually building the thing, and somewhere in there a designer would push back, or an engineer would ask why, or a new component was needed for the design system, or it would fall apart on contact with a use case that wasn’t Slack’s. In short, the friction was doing quality control.

Now a request to make it look more like Slack is returned before you finish your morning coffee. You prompt it, and there’s no one in the loop long enough to object. Slack is Slack partly because of what they wouldn’t build. You can copy their nos. You can’t copy their judgment.

The model doesn’t know what room it’s in. It can’t tell if the black keys are broken, or whether anyone in the audience is enjoying what’s being played. It plays like a piano roll: every note in the right place, and no idea anyone is in the room. We keep handing it work that needs somebody to take a position.

We built design systems for each other. Humans making things for humans, with documentation about how and when to use the patterns, which usually ran a few sprints behind. They were great for things we knew about, but they weren’t great at what to do if things went wrong. Sometimes they included a do-this column and a don’t-do-that column. A designer read the doc, saw what it didn’t say, and filled the gap with judgment. The system was never the artifact. It was the artifact plus a human interpreting it.

AI doesn’t work like that. It ingests the system and needs the if-this-then-thats worked out in advance, because while it will exhaust tokens researching what to do next, it will not stop and wonder what we meant the same way a human does. Maybe it gets there. If it can crack hundred-year-old math problems, it can probably enumerate every permutation of every failure. But the system only holds answers to questions somebody thought to ask. It can tell you a flow violates a rule. It can’t tell you a flow nobody has ever complained about is wrong.

Design systems as we currently think of them, the border radius variables, color tokens, component libraries, and documentation, will likely be upgraded to speak to machines first and humans second. The next phase is working out the logic for those if-this-then-thats, in .md files, loops and MCP servers that agents can ingest. That work is in progress. I’m on my fourth iteration of a Claude workflow to optimize for this, and it’s worth doing. It will get you to about a 7 out of 10.

For now, that’s where AI tops out reliably. It’s also where the market often says “good enough.” A 7 ships. A 7 tests fine. Nobody gets fired for a 7, and the last three points now look like headcount you can cut.

When Jarrett found out the piano was flawed, he could have coasted. Reached only for the phrases that had worked before, capped the concert at forty-five minutes of improvisation assembled out of moves he already had. Nobody in that hall would have known.

That’s the 7, and it’s what a model does every time. It generates. Fresh, unrepeated, and made completely of things that have worked before. The median, delivered confidently.

At the 25-minute mark, Jarrett has just spent five minutes coming out of a peaceful section into an intense build, and there doesn’t seem to be a way out. He pulls an abrupt phrasing out of nowhere, ending Part 1, and the audience claps.

The applause is the 7. It worked, the room is happy, and this is where almost everyone stops. He waits for them to finish, then plays four notes, because he wants to know what happens if he tries again.

You are going to get the wrong piano. The system will be half built, the research will have been cut, the roadmap will be somebody else’s, and the model will hand you a hundred versions that are all technically correct.

You will fix it. It will be pretty good. People will tell you so, and that is your applause. Nobody is going to ask what comes after it, so you have to: is a 7 good enough here, or is this the one worth the last three points? Sometimes the 7 is the right answer. Knowing which one you’re in is the job.

When it isn’t, the push pays three ways: the work gets better, you get better, and eventually the model does too. Play what the room gave you. Keep going through the part that isn’t working yet. What has to be decided in the room is the part that was always yours.