Ideas

Working through the interface

I’ve been exploring a model-agnostic canvas, with the workspace itself as the thing that stays consistent. That leaves room to try different models without making each one a separate place to work.

The current frontend lets me zoom around the canvas, focus on an item and interact with example widgets, including a timer. For a portable demo, I used local examples so I could inspect the canvas interactions independently of live generation. That demo leaves out AI calls, Notion and backend storage.

I also tried bringing it into Figma as an interactive Code Layer. The import was blocked by the tool, so the working version is still the browser demo.

The open question is what makes these generated interfaces worth keeping and returning to. That’s what I want to explore beyond the initial moment of generating something.

Input is an experiment in giving agents somewhere to ask for decisions. A thread can send a card with a proposal, options or an open question. I can take action, give direction or dismiss it, and the answer goes back to the conversation that asked.

Trying it on itself

I started with small test cards to check that choices and comments made the round trip. Then I used Input to guide its own development: asking the agent to feed me decisions about what to work on next, rather than managing everything through a parallel chat.

That made the response as important as the question. The agent needed to show that it had received my answer and record what happened next. Choosing a product direction also needed to stay separate from approving an implementation.

Using it across desktop and mobile led to smaller UI changes too: keeping the tabs accessible, matching the header to the page and asking for a quiet arrival sound without another visible control.

The next question is how other agents could use the same inbox. The current experiment has been dogfooded through ChatGPT.