Chat is the least interesting interface to LLMs

Large language models are an amazing tool for working with prose inputs and outputs (with chat interfaces being the prototypical example). In fact, they are so good at working with prose that advocates often forget that prose is actually a pretty crappy user interface on almost every other dimension:

• users tend to prefer structured inputs (e.g. buttons, sliders, date pickers) over writing large amounts of prose

• users tend to prefer structured outputs (e.g. graphs, tables, diagrams) over reading large amounts of prose

• prose is poorly suited to mobile devices (where it’s painful to read/write long passages)

• prose communication is inherently ambiguous, like cutting with a dull knife

None of these limitations of prose are inherent limitations of LLMs, though:

• LLMs can generate structured outputs

• LLMs can accept structured inputs

• LLMs can even generate structured form controls for the user to drive them

  See, for example, the popup-mcp tool or Grace

It’s just that people don’t lean heavily enough on these options and they should! When a user interacts with a model (or agent) you want to offboard them off of prose and onboard them onto a structured interface as quickly as possible. Not only will this improve the user experience but it also improves user confidence: a button labeled with a clear call to action promotes confidence in a way that a freeform text input does not.