The "Gulf of Execution" in AI chat is real - and it’s exhausting.
Staring at a blank prompt box, trying to engineer the perfect text, and back-and-forth messaging just to tweak a setting? That’s not productivity. That’s friction.
Chatting about work isn't the same as doing work.
The fix? Generative UI.
Instead of forcing users to type everything out, AI should build the interface on the fly:
Card-based UI: Quick, scannable micro-actions over walls of text.
Auto-generated forms: Pre-fill what the AI knows; let the user tweak the rest in two clicks.
Custom applets: Drop mini interactive tools right into the thread when visuals or calculators matter.
If I ask an AI to configure a project, compare several options, review a diff, or choose from a tree of files, why should it respond with another wall of text?
The model can decide to create a form, table, selector, confirmation dialog, diff view, etc. — without being specifically trained for GUI generation.
The UI is retained locally, so sliders, checkboxes, typing, and other low-level interaction don’t require another LLM call. Only meaningful events such as Submit, Confirm, or Explain go back to the model.
So the flow becomes:
LLM → UI tools → interactive UI → semantic event → LLM
instead of endless:
LLM → text → user → text → LLM
It feels a little like “Dear ImGui for AI agents.”
Open standards like A2UI (Agent-to-User Interface) are making this seamless - letting agents output clean UI components directly into any app, not just text strings.
The next era of AI UX isn't teaching people to prompt better. It's building interfaces that get out of the way.
Are you still building pure chat, or moving toward Generative UI?
I for one built a small PoC library allowing to expose custom UIs to ChatGPT and similar AIs. Imagine running the whole CAD or 3d slicer or video editor INSIDE the chat window.
You can even play Pong while boss is not looking!
Here's the link to the library: Agent UI
Pong:
Configure CMake UI:
Top comments (0)