how Shoti became a night city with a voice, and why chat was never the destination
Shoti’s interface is a city at night, with a narrator who tells you the story of the work. That sounds like a flourish. This essay is the argument that it is the opposite of one: the interface between people and machine-speed work is not chat and not text, but the two channels humans were built for. A scene you can glance at. A story someone tells you.
When Grace Hopper proposed a programming language made of English words, she was told, in her own recounting, that it would not work because computers did not understand English. She shipped it anyway, in 1956. When IBM released FORTRAN the next year, veteran programmers dismissed it as a toy; within about a year, more than half the code running on IBM’s machines came out of the compiler. The pattern has repeated every decade since: each new level looks like cheating to the people one level down, then it becomes the floor. Almost nobody writes assembly today. Assembly runs everything.
The newest level arrived on schedule. “The hottest new programming language is English,” Andrej Karpathy wrote in 2023, and by 2025 he had named the practice of surrendering to it entirely: vibe coding, where you “forget that the code even exists.” You describe intent in your own words and the machine handles every level below.
Each step, from Hopper to Karpathy, raised the level of the language you write. The reading side got tools instead of new languages: debuggers, profilers, type checkers, review interfaces, observability dashboards, each one a better window onto the same low-level artifact. You still read code, diffs, logs, and stack traces. The window got clearer. The language behind it never changed. And the windows were sized for a person’s output; a fleet of agents overflows them. What comes back from an agent is markdown files, diffs, logs, walls of bullet points: machine-format text, at machine volume. The format is honest about its audience now. There is a standard called llms.txt, in which websites publish clean markdown specifically for language models to read (this site publishes one). Markdown was designed for people to write; its heaviest readers now are machines. Most of what your agents send back is the same shape, structured for parsing rather than comprehension. It only looks like it was written for you because it is in English.
You now write in English, and you still read in assembly. In the first essay we showed the review queue backing up as the serial human step became the whole job. This is the same finding seen from inside one person’s day. The bottleneck moved from producing the work to comprehending it, and comprehension never got a higher-level language.
The industry’s reigning interface to all of this is a text box, and the text box is not there because anyone concluded it was right. It shipped because it was easy to ship. The critiques have been on the record for years: a bare prompt box has no affordances, Amelia Wattenberger’s point, a rectangle that could equally be a search field or a wishing well. Daniel De Laney aimed the sharper version at coding tools: “Chat reduces you to memory and hope,” and “You don’t program by chatting. You program by writing documents.” Maggie Appleton says the quiet part about the form itself: the general-purpose assistant box is a transitional interface, the shape a technology takes before each domain grows the interface it deserves.
None of this means chat dies. The terminal never died either. Chat is this era’s terminal: the raw layer underneath everything, always available, the right tool for the person who wants to touch the machine directly. What we refuse is the idea that a team should live there, forwarding each other walls of agent transcript. The editors and terminals are magnificent at the hour of writing code. Our problem is the other twenty-three hours: knowing what all of it means, across every person and every tool, and what to do next. We are not trying to out-chat anyone. We are trying to out-understand them.
During the map redesign I kept repeating one sentence at my own prototypes: we prove physics by calculations, but humans shouldn’t compute. Give them something they can consume. It became the test every version had to pass: a person who has never seen Shoti looks at the screen for five seconds, with no legend, and whatever they can say afterward is everything the screen communicates.
Be careful with the folk version of this idea. “People are visual learners” is a myth; the learning-styles theory it comes from has failed its tests for two decades. What is true is narrower and stronger. The visual system is fast in a way reading is not: the brain classifies a natural scene in something like 150 milliseconds, and in Lionel Standing’s 1973 experiments people shown ten thousand photographs could still pick out, days later, which ones they had seen. Larkin and Simon’s classic paper on why a diagram is sometimes worth ten thousand words located the advantage precisely: a good diagram indexes information by location, so the reader searches less and sees more. The qualifier is theirs, and it is the design constraint. A diagram beats text when its geometry means something. When it does not, it is decoration with a legend.
Our own first maps failed this test: we drew the company as abstract islands, then as sectors of a circle, and both were geometry that meant nothing without a codebook. The city works because you already own the codebook. Lit windows at night mean someone is in there working. A figure walking the street means something is arriving. A building under construction means something is being built. Two crews converging on one site means trouble, and that one is the catch, standing in the middle of the scene it caught. Names and numbers stay in words, in the labels and the one composed sentence above the city, because precision needs words. Any channel that needs a legend is banned. We compute so the human never does.
A glance tells you where things stand. It cannot tell you what happened while you were away, why it matters, or what to do about it. For that, humans have always used the same technology: someone who knows tells you the story. The newest thing in Shoti is that someone.
Ask the city anything, or just come back after a day away, and the narrator tells you the story of the work: a composed brief with a beginning, a mechanism, and a consequence. Under the hood a story is a list of beats, and a beat is a sentence, a camera move, what lights up, and a next step. While the voice speaks, the camera flies; the words and the scene are one channel, not two competing ones. Every story ends with something to do rather than a dead end: build from here, fly there, hear the deeper story, see who to talk to.
Two rules keep this from being a gimmick. First, the narrator is compiled, never invented. Every sentence must trace to the record the team’s work already emits, the same atoms the catch runs on: intent, scope, actor, time. One fabricated sentence about your own company would kill the narrator, so on thin evidence it says less rather than making more up. Second, the story is how work starts. The compilation the narrator gathered to tell you the story is, unchanged, the context packet an agent needs to start the work correctly. In the shipped product that is a click: open your codebase as a city, point at a place, and your agent starts there, with the real files, the ranked symbols, and an honest account of what it is not sure about, instead of a blank chat box. Blank-box prompting is what Shoti replaces. Understanding and starting are one motion.
Why trust something this old? Douglas Hofstadter’s Stanford lecture “Analogy as the Core of Cognition” gave me the frame. Language, he argues, is “a drastic compression of the complex symbolic dance occurring in the sender’s brain,” powdered food for thought; “what matters is not the dried linguistic powder that is used to transport the dance between brains,” what matters is the dance it sets off in yours. A few black marks on a white background and you see, in his example, a fishing village, a pier, a figure pacing at the end of it. A transcript dump is powder shipped by the ton with no thought for the reconstitution. A story is powder measured for one mind, in one moment, and it lands: in Bower and Clark’s experiments, people who wove word lists into stories recalled 93 percent against the memorizers’ 13. The scientific paper and the formal proof are the recent exception, built for specialists verifying each other. Most of a company, most of the day, is not verifying. It is trying to understand.
We did not set out to build an interface. We set out to stop teams from wasting days on work already done, already killed, or about to collide, and the case for that is the first essay. For months we treated it as two problems: catch the wrong work, and separately, make the work legible. Stepping back to solve the second dissolved the boundary. The atom the catch already ran on, intent, scope, actor, time, is the same atom that places a building on the map and gives the narrator its next sentence. One signal warns your teammate, lights the city, and becomes the story. Nothing extra is collected, and nothing shown was ever written by hand. Coordination stopped being a feature we bolt onto an interface; it is what telling the true story of the work, in time to act on it, does on its own.
The size of the answer is the part I trust most. Knowledge has always grown the way Popper described, by conjecture and refutation, guesses criticized until something survives, and David Deutsch gave the test for what survives: a good explanation is “hard to vary,” and its nature is “that there is only one of them.” Hindsight is how you recognize one. When Huxley finished Darwin’s argument, his reaction was a groan: “How extremely stupid not to have thought of that.” The constraint that makes this design hard to vary is the record. Split the warning, the map, and the story into separate surfaces and each needs its own feed. Any feed not emitted as a byproduct of the work has to be written and kept current by someone, and that upkeep is the tax the whole design exists to remove. One record is the only version where nothing has to be manufactured. Now it looks like the only move. It was invisible for months, and every version we killed was a way of not yet seeing it.
Why a voice at all? Because the moments that need narration are the moments your eyes are busy. Aviation settled this decades ago: when an aircraft is near the ground, the plane speaks, “pull up,” “sink rate,” the altitude callouts, because a pilot in that moment cannot spare a glance for one more instrument. A Formula 1 engineer narrates the car’s state to a driver who cannot look at a dashboard. A screen reader delivers an entire interface as spoken state. Voice is what you use when attention is somewhere else and needs to stay there.
The discipline for doing that without becoming noise was written in 1996. Mark Weiser and John Seely Brown called it calm technology: the good tool “engages both the center and the periphery of our attention, and in fact moves back and forth between the two,” where the periphery is “what we are attuned to without attending to explicitly.” Amber Case turned that sketch into engineering: eight principles, the first of which is that technology should require the smallest possible amount of attention. Shoti is built to that reading of the world. The city at night is the periphery; you are attuned to it the way you are attuned to weather through a window. The voice is the licensed trip to the center, and it makes that trip on invitation. The brief offers play or skip. The founding tour plays once. The catch and every alert never speak; the voice explains, and it never alarms. And the voice runs on your machine, a small model of about eighty megabytes, one cast voice, no cloud in the loop, so the story of your company is told inside your company.
In March 2026, Paul Graham published “The Brand Age,” a history of what happened to Swiss watches after quartz made mechanical excellence irrelevant. His definition is the useful instrument: “Brand is what’s left when the substantive differences between products disappear.” When a field’s real problems get solved or stop mattering, products converge, and brand floods into the vacuum; watches grew larger and louder because there was nothing left to compete on but recognizability. His verdict on that phase is blunt: stay away from brand. The golden ages belong to the people following interesting problems while the problems are still open. He is describing watches. He is also describing a possible future of AI tools, most of which are already converging on the same rectangle with the same text box, exactly the condition under which brand takes over from substance.
I bring this up because a night city with a narrator will read, on a screenshot, as brand: the distinctive-looking one. Graham draws the line that matters here: “Branding by definition has to be distinctive. But good design, like math or science, seeks the right answer, and right answers tend to converge.” So it matters to us that every choice in the city traces to a constraint rather than to a desire to look different. The scene channel comes from what a glance can carry. The words carry precision because scenes cannot. The voice exists because eyes are busy, waits because attention is expensive, and runs on-device because the story of your company is yours. Graham’s essay has a warning label for companies like ours, the slow drift he calls the comb-over, small steps toward decoration that no one notices from inside. Our tripwire is simple to state and will be hard to obey: the day we make the city prettier instead of truer, we have started combing. Coordinating many people’s agents, and making machine-speed work comprehensible to the humans responsible for it, are unsolved problems. That is the position. Follow the problem while it is still a problem.
The first essay ended with a list of what we could not prove, and the habit is worth keeping.
Chat will still be there, the way the terminal is still there, and some days you will want it. The bet is that you will live above it. So, one check: of everything your agents wrote yesterday, how much did you read? You already write in the highest-level language. You should get to read in one.