Every time you send a message in a chat app, you follow the same ritual: open the app, find the contact, tap the text field, type, hit send. Five steps for one intention. A developer named Nand has published a framework called UI2 - Unified Intent Interface - arguing that AI should collapse all of that into a single natural language command.
The concept isn't new in broad strokes. Apple Intelligence already does some version of this, and every AI assistant promises to "just do what you ask." But the UI2 framework is more specific about how it should work: the user states an intent in plain language, an LLM (large language model - the AI architecture behind tools like ChatGPT and Claude) translates that intent into structured commands, and the app executes them through a defined schema.
The key difference from current AI assistants is the feedback loop. Today, when you tell Siri or Google Assistant to do something, it either works or it doesn't - you find out after the fact. UI2 proposes a confirmation step where users verify the interpreted action before it runs. Tell the system "message Sarah that I'll be late" and it shows you the parsed intent - recipient: Sarah, message: I'll be late, app: Messages - before executing.
The author predicts more apps will add AI-powered command palettes that map natural language to available actions. We're already seeing early versions of this. Notion's AI can create pages and databases from descriptions. Linear lets you create issues with natural language. Raycast has built an entire launcher around this pattern.
But there's a tension the framework doesn't fully resolve: intent-based interfaces work best when the app has a finite set of actions. Sending a message has clear parameters. But "make this spreadsheet look better" or "organize my project" are ambiguous in ways that a schema can't easily capture. The more complex the task, the more the natural language interface starts to feel like a conversation rather than a command - and conversations need a different UI pattern entirely.
The most practical near-term application is probably what the framework calls "task stacking" - chaining multiple simple actions into one request. "Email the team the Q2 report and schedule a review meeting for Thursday" touches two apps and four actions, but expresses a single intent. That's where the friction savings are real, and where current interfaces fail hardest.