Yesterday's News

A short essay on something that happened the day before.

Fish Cakes and Promptecture

Prompts can produce extraordinary images of architecture. But what happens when we give up some of the constraints that make architecture what it is?

In yesterday’s ArchDaily, architect Romullo Baratto notes that architecture now often begins with words rather than images.1 His example is a prompt for a cultural centre on a desert coastline: »in the style of Alvar Aalto, wood structure, golden hour, aerial drone shot, photorealistic, 4K«. That’s an image brief. But is it an architectural design?

»Such images expand our understanding of architectural potential. They scaffold our dreams, sometimes enough to influence what gets built«

Architecture has always used words, drawings, and models to imagine things that don’t yet exist. What’s in the drawing doesn’t have to be buildable. It can represent an impossible city, a space without gravity, a building made from shadows. Even speculative drawings give a form that can be examined, developed, or rejected. Such images expand our understanding of architectural potential. They scaffold our dreams, sometimes enough to influence what gets built.

But where a drawing narrows possibilities, a prompted image does something different. It shows a building’s appearance precisely while leaving almost everything about its making undetermined. Let’s call this promptecture – producing architectural possibilities by deliberately relaxing the constraints through which architecture acquires form.

There’s something uncanny about this. We recognise our intentions in the image, but can’t quite account for what’s appeared. The model fills in what we left unsaid, drawing on patterns learned from vast collections of images. The image is ours and also not quite ours. It surprises us precisely because we surrendered some control over what it will become.

Philosopher Vilém Flusser helps us understand this. In Towards a Philosophy of Photography, he describes how technical images translate concepts into pictures through the programmes of the apparatuses that produce them.2 An image isn’t a transparent window onto an idea; it’s shaped by the system through which that idea becomes visible. And a prompt isn’t neutral. »Golden hour«, »aerial drone shot«, and »photorealistic« already encode conventions about how architecture should be seen. The model interprets our descriptions through a visual culture that precedes both the prompt and its author.

Theorist Mario Carpo has traced a related shift in architectural computation: from systems built on inspectable rules to those whose workings are opaque.3 Architect and theorist Christopher Alexander’s A Pattern Language tried to make spatial knowledge explicit, a shared vocabulary anyone could use to generate form.4 A prompt may look similar, but it operates through statistical associations rather than readable patterns. We can describe what we want without knowing why the model produces what it does.

That uncertainty isn’t necessarily a problem. Generative systems can easily produce thousands of variations in the time it takes to draw a single proposal. They can take us somewhere we wouldn’t have thought to go. The question is whether this freedom expands our imagination or simply reproduces familiar architectural imagery.

An intriguing precedent is artist Tobias Rehberger’s Tod Man Plaa (»Fish Cake«), made in 2004. He commissioned craftsmen in Thailand to build full-scale replicas of familiar cars, giving them sketches and verbal descriptions, but no technical drawings or measurements. For a Renault Alpine, he provided memories of the car over the telephone, without sending any pictures. The replicas had to be human-sized and drivable.5

The craftsmen had to interpret the instructions, supplying everything that had been left unsaid. Their knowledge, assumptions, and practical judgement became part of the translation. The differences between the originals and their replicas weren’t mistakes. They were the work itself: the consequences of an incomplete description passing through another intelligence.

Generative models similarly fill in what a prompt leaves open, but through statistical relationships learned during training rather than an understanding of how something works. But Rehberger’s craftsmen had to make their interpretations answer to the world. The cars had to work.

Architecture gives form by imposing constraints. Promptecture finds possibilities by relaxing them. The two aren’t enemies. A prompted image can be the start of an extraordinary building, just as a drawing of an impossible city may change how we think about real ones. But it doesn’t have to resolve the challenges it creates.

The image can remain an invitation. A building becomes an obligation. Promptecture can take us anywhere. Architecture has to bring us back to earth.

1 Romullo Baratto (2026) »Beyond Buildings: ArchDaily’s November Editorial Focus«, ArchDaily, 3 November 2026. https://www.archdaily.com/1187895/beyond-buildings-archdailys-november-editorial-focus

2 Vilém Flusser (2000/1983) Towards a Philosophy of Photography, translated by Anthony Mathews (London: Reaktion Books; originally published in German in 1983).

3 Mario Carpo (2011) The Alphabet and the Algorithm. Cambridge, MA: MIT Press; Mario Carpo (2017) The Second Digital Turn: Design Beyond Intelligence. Cambridge, MA: MIT Press.

4 Christopher Alexander, Sara Ishikawa, & Murray Silverstein (1977) A Pattern Language: Towns, Buildings, Construction. New York: Oxford University Press.

5 Tobias Rehberger (2004) Tod Man Plaa. See LABoral Centro de Arte, project description, https://laboralcentrodearte.org/en/artworks/tod-man-plaa-2004-2/; »Tobias Rehberger’s Cars«, We Make Money Not Art, 26 May 2009, https://we-make-money-not-art.com/tobias_rehberger_tod_man_plaa/; and »Tobias Rehberger’s Thai-built Cars«, Jalopnik, 21 March 2013, https://www.jalopnik.com/these-amazing-thai-knockoff-supercars-might-be-the-grea-456756306/.


Previous editions

See the full archive →

Subscribe

Get it daily, or just the weekly roundup.