Should an AI tool's design follow the type of task it does, or the stage you're at in using it?

This explores whether AI interface design is better organized as a catalog of pattern types or as a sequence of behavioral phases a user moves through. No note in the collection compares the two head-to-head. It does contain clear examples of each approach, and reading them side by side shows what each one buys you.

The phase-based approach is clearest in Friedman's argument that chat makes users spell out what they want slowly, one typed prompt at a time Can task-oriented UI patterns reduce the friction of AI chat interfaces?. Friedman's five design levers read almost like a timeline: capture intent, shape the output, refine it, delegate agentic tasks, embed the result in a workflow. Organizing by phase keeps the designer focused on where users actually get stuck. In Friedman's case that is the very first step, where the user has to put intent into words. The pattern-type approach shows up in two taxonomies. The IAI model looked at 66 tools and reduced them to twelve atomic interaction paradigms that can be combined How do interaction signals reshape what generative AI actually receives?. A review of 291 papers mapped the design space along six dimensions, including user control, model feedback, and engagement level How should designers structure human-GenAI interaction systems?. The pattern catalogs make it easier to compare tools and reuse parts. The IAI authors claim their model allows design comparisons that earlier taxonomies could not.

The surprising twist is that the best pattern taxonomy is secretly about something else. The IAI model's key move is to ask what the AI actually *receives*: the prompt plus all the constraints the interface adds to it. That is neither a pattern list nor a phase sequence. It is about how information flows. The same idea appears in the argument that AI context keeps shifting (prompt, history, retrieved data) in ways users can't keep track of the way they learn a fixed menu How does AI context differ from conventional software context?. If context changes at every turn, a static catalog of patterns misses what matters, and a neat set of phases can blur together too. Lam and colleagues push a related idea. Their design surfaces the 'generate, then evaluate' decision at every interaction instead of fixing it once up front Does visible design structure keep HCI expertise necessary as AI advances?. In effect, every moment becomes its own small design phase.

An older lens from outside interface design offers a third answer. Malone and Crowston's coordination theory says to organize around *dependencies*: first work out how activities depend on each other, then pick the coordination mechanism that fits Can we match dependency types to coordination processes?. Applied here, you would first ask what each step in the user's journey depends on (a clear intent, a reviewable output, trust in the agent), and only then choose a pattern. Dubberly's agent stack points the same way with two axes at once: capability layers and oversight scale How should agent capabilities be organized into functional layers?.

The practical takeaway is that this is probably a false choice. Phases tell you where the friction is, and patterns tell you what tools exist to reduce it. The corpus suggests a third organizing idea that may be underused: track what context the AI holds and what the user can see at each step. Generative UI tools offer a warning about what goes wrong otherwise. They skip about a quarter of the design reasons they're given and recognize only half of the UX principles in their prompts Do generative UI tools actually implement their stated design rationales?. Whichever way designers organize their toolkit, the reasoning behind it needs to stay visible, or it gets lost.