The most consequential design deliverable I've shipped recently contains no screens. It's a set of rules, safety rails and evaluation criteria, and it governs what an AI system is allowed to generate.
Generative interfaces change the basic unit of design work. When the system composes the surface at run time, shaped by context, history and intent, there's no fixed screen to perfect. What persists is the envelope the generation happens inside. Someone has to design that envelope. If Product Design doesn't, an engineer's defaults will, and a model's defaults aren't a design position.
i.What a constraint system contains
In practice it has three parts. Rules describe what the system may do: the components it can compose, the tone it can take, the actions it can offer. Rails describe what it must never do, the surfaces it can't generate, the claims it can't make, the data it can't expose, the moment it has to hand over to a human. Evaluation criteria describe how we know the output is any good: the checks a generated surface has to pass before a user sees it, and the signals that tell us when the envelope itself needs to move.
This isn't hypothetical paperwork. The rules become prompts and schemas. The rails become refusals and fallbacks. The evaluations become tests that run against generated output the way unit tests run against code. Design review doesn't disappear. It moves from inspecting instances to inspecting the system that produces them.
ii.Provenance is a design requirement
The other half of the discipline is transparency. When an AI surfaces a suggestion, the interface has to show where it came from and why. People don't adopt what they can't interrogate, and they're right not to. Provenance isn't a tooltip bolted on at the end. It's a constraint written into the system from the start, carrying the same weight as a safety rail. To the user, a suggestion with no visible source is indistinguishable from a guess.
The same goes for personalisation. If a system is adapting to someone, they should be able to see what it knows, why it's adapting, and what's being passed onward. Inferred data deserves the same care as declared data, because the user experiences both as claims about who they are.
iii.The craft doesn't shrink, it moves
Designers sometimes hear all this and grieve the loss of the canvas. I understand the feeling and think it's misplaced. Writing a constraint system well demands exactly the craft we say we value: precision, a feel for how people will actually meet the mechanism, and the humility to encode your judgement clearly enough that a machine can apply it when you're not in the room. The canvas got bigger. It just stopped being a rectangle. ✳