UX Collective: The prompt is not an interface
What the article is about
Joshua Leigh’s May 2026 essay for UX Collective makes a structural argument about AI and interface design: the current wave of AI tools has sent users back to typed commands, which is a step backward from the decades of work that established direct manipulation as the basis of usable software. The article does not dispute that AI-assisted design tools produce output faster. The argument is about how the interaction model has changed and what that costs designers working with it.
Context and argument
The essay opens by observing that the prompt-based interaction model used by most current AI tools is essentially a command line. Natural language replaces exact syntax, but the user is still formulating intent in text and waiting for a system to interpret and execute it. This is a fundamentally different model from the visual feedback loops that tools like Figma established — where changes appear on the canvas as you make them, and the designer develops a feel for the material in real time.
Leigh identifies three specific problems with the prompt model for design work.
The first is the loss of visual intent. Designers think spatially and compositionally, not primarily through language. Translating a layout intuition into words is a lossy process. What the text describes and what the designer imagines are not the same thing, and the gap often requires several generate-evaluate-revise cycles to close.
The second is the loss of direct manipulation. The ability to drag, resize, and visually constrain elements gives designers immediate feedback on whether something is working. Prompting replaces this loop with a cycle that is slower and less embodied.
The third is aesthetic underdifferentiation. Because AI tools generate from training data, outputs tend toward the mean of what the model has seen. The result is work that is recognizable as AI-generated — not because it is low quality, but because it lacks the decisions that reflect a particular designer’s choices.
What it does not cover
The article does not propose a fully formed alternative interaction model, though it suggests that the most effective AI design tools will be those that combine generation with direct manipulation rather than replacing one with the other. It does not compare specific tools.
Who it is useful for
This piece is worth reading for designers who have worked with AI tools regularly and noticed a discomfort they have not yet named, and for product designers and UX researchers who think about how interaction models shape creative output. It is also relevant for people building design software who are deciding how AI generation and manual editing should coexist.