AI UI Sandbox

An ongoing batch of inline UI sketches — the agent answering with a small working module instead of another paragraph of prose.

When communicating the order of info would be harder with the standard question-and-answer UI.

Also my excuse to make a better-feeling drag-and-dropper. Most mobile drag-and-drops feel clunky.

Chats with agents get long, and the part you need to get back to is often buried. Instead of rereading it all, the app could help by post-processing the chat transcript.

Find the part in the chat you half-remember, plus context labels for near-matches.

UI details for when you cut it off mid-answer.

When the agent isn't sure what you meant, it can show the readings it's considering and let you point, instead of spending a turn asking.

Micro-animation studies for agent text as it arrives.

A reply that leans on an earlier turn carries the quote with it, so you can check the source without going to find it.

Stored memories usually all look equally fresh, however old they are. Here a memory wears visibly as it ages — a cue to re-confirm before it gets used.

An ambient sidebar readout of how full the current conversation's context window is — so the diminishing-returns zone is visible before you hit it. Could roll in token cost too, or stay purely about context budget.

How an agent looks while it works — retrieving, deliberating, hesitating, etc.

Quick ways to confirm what an agent did without redoing the work yourself.

Where a claim came from, attached to the phrase.

Hold-to-confirm gestures keyed to how reversible the action is.