5 · PROMPT-SHAPED ARCHITECTURE

THE CITY

Picture a city built entirely by day-crews. Each morning a crew arrives at the gate with no memory of the place and no knowledge of any other crew. The mayor tells them what is needed. They build it — competently, quickly, exactly to the order — and at dusk they leave and never come back. Tomorrow a different crew arrives, equally skilled, equally blank.

Every building will stand. Every one will be built well. And the city will come out shaped like the conversations, not like the place.

Walk one street and you will find four warehouses, each with its own measuring stick, cut from scratch by four crews who each had to invent a measuring stick because nobody told them one already hung in the town hall. Three older warehouses across the street use the official stick. No one decided that. No one could have decided it — there was never a moment when a person who knew about both stood in one place.

This is the quiet disease. It breaks nothing today. What it does is write the transcript of how the city was ordered into the city's own bones, and then bill you for it every single time you want to change one thing everywhere.

THE TELL

• You ask for one change "everywhere," and you get it in some places — and nobody can tell you in advance which places.
• The same job is done well in one part of your product and badly in another, and you never chose the difference.
• You can date a feature by its shape rather than by what it does.
• A fix lands in one district and never reaches its identical twin two streets over — and the twin's existence was news to everyone.
• You ask "how does the city do X," and the answer starts with "which one?"

FROM THE FIELD

The field record is thin here — this chapter argues from the mechanism, not from a scar. The first engagement produced exactly one receipt for this pathology, and the fit is contestable; eight further deep reads of the city's own decision record turned up nothing more. That is the honest state of the evidence, and padding it would make this the one chapter in the book you could not check.

One more honesty is owed before the receipt. The field manual behind this survey files "functions written in different sessions, neither aware of the other" under Pathology 1, Parallel Systems — which means this disease is routinely arrested under its sibling's name. Whether Prompt-Shaped Architecture and Parallel Systems are one condition or two is an open question inside the instrument itself, not a settled matter we are reporting to you (Engagement 01, R5.2).

Four sticks, one town hall

In the first city, seven of its translators each took an outside industry data feed and turned it into the city's own records. Four of those translators carried a byte-identical copy of the same small piece of logic — a reader for calendar dates. The copy could not read a date format those feeds commonly used. The other three translators, sitting in the same archive the whole time, used the city's correct shared version, and always had.

The ledger's own summary of what went wrong — "mappers" being its word for these translators: "The knowledge never crossed between mappers." (the first city's ledger, DL-064)

What makes this the disease rather than plain duplication is how the four came to exist. Each one reads as the output of a single build session, blind to its siblings. The pattern is not one worker copying its own work four times over; it reads as four separate shifts, each solving the same problem from first principles, none of them aware that three doors down the correct answer had been sitting on the shelf the whole time. The shape of the street records the sequence of the conversations. Nothing else explains why four went wrong while three, in the same archive, never did.

WHY AI WORKERS DO THIS

Two properties of the worker do all the work here, and neither is mysterious.

The first is that it starts every shift with no memory. The worker that wrote the fourth date reader had never seen the first three, and nothing about its shift made it go looking. The second is that it never refuses. You asked for a reader, so you got a reader — the question on the table was "write me one," not "do we already have one," and a worker that cannot refuse also cannot redirect you.

A human builder carries the city around as a nagging weight. That weight is exactly what makes a person stop and say hang on, don't we already have this? The AI worker carries nothing. Its entire world is the request in front of it, so the request becomes the design. Which means your architecture ends up shaped like your own request history: careful where you were careful, duplicated where you asked twice, arbitrary where you asked casually, and coherent only in the stretch where you happened to explain the whole picture in one sitting.

THE COUNTERMEASURE

"Don't duplicate work" is not a law a worker with no memory can obey. The law has to put the answer somewhere the next blank crew will trip over it before they start swinging.

Give your workers this, verbatim:

Before you write anything that computes, formats, reads, or decides, name the thing in this city that already does that job — by its address. If nothing does, state in writing that you looked and where you looked. Anything used in more than one place has exactly one home, and that home is listed in the city's index, which every shift reads first.

The second half is enforcement, and it is yours, not theirs: when you fix something of this shape, never accept a fix in one location. The order is find every other place with this shape, list them for me, and fix the class.

This narrows the disease. It does not end it. A worker can search honestly and still miss, because it searches by name and the twin was named something else. The index is the control; the searching is only a habit — and habits are what drift first.

ASK YOUR OWN CITY

1. Ask two different parts of your product to name the same thing. Do they use the same word? Do your builders use the same word?
2. Go looking — do not guess — for two places in your city doing the same job: two screens computing the same number by different rules, two helpers that almost match. Where are they?
3. Does the newest part of your city follow the same conventions as the oldest? Find the moment the convention changed, then find out whether anybody decided it or it simply happened.
4. Does one thing your users touch — a table, a form, a date picker — exist in three separate versions because three shifts each built one?
5. Ask for your whole product drawn on one page — every part, and which parts touch which. If nobody can draw it, that is the finding: the shape exists, and no one is holding it.