Ask an assistant for a small interactive tool and it builds one inside the conversation, then saves it as an app that reopens with the state it had when you left it. Requires an assistant with a toolchain agent, which is the only type that builds apps, added in Add and configure agents.
1
Describe the tool you want
Open the assistant and describe what the tool does, rather than how to build it. There is no need to mention frameworks, components, or storage, since the agent picks those.
a request
"A workout logger: pick an exercise, log sets and reps, and show last week's numbers for that exercise underneath."
It already knows your material
The agent can read the assistant's memory notes, so an app built by a recipe assistant arrives populated with those recipes. You do not need to paste your data into the request.
2
Revise it in the same conversation
Describe the change you want and it edits the app that is already there.
revision requests
"Make the completed ones collapse." "Sort by due date instead of priority." "Bigger tap targets, I use this on my phone." "Add a reset button that clears the week."
Asking again for the whole app in one sentence replaces it rather than editing it, which is how to start over.
3
Ask for what it should remember
Name what should persist, since an app keeps only the state you ask for and otherwise resets when it closes.
asking for state
"Keep a running total." "Remember which ones are checked." "Keep a log of entries with the date."
4
Save it
Use Save app. The dialog takes a Name, an optional Description, and a State scope that defaults to Device.
state scope
Scope
Who shares the state
Device
Each device saves state
User
Each account saves state
Shared
Everyone shares one live state
Change the scope later by editing the saved app. Choosing between them is covered in Keep state in an app.
Why a saved app reopens instantly
Building an app calls a model. Saving stores the finished program, and opening it afterwards runs that stored program without calling a model again, so a saved app opens immediately however often you use it.
Saved apps are listed in Saved apps, where each one can be opened, shared, or published as a page.
If a build fails
Generated apps are checked and run before they reach you. When something is wrong, the error goes back to the agent and it tries again a few times, so a failure message means that loop was exhausted. Asking again works, and so does asking for a smaller version first and adding to it.
What apps can and cannot do
Apps can hold state, respond to input, draw interfaces, and calculate over your data. They cannot reach the network, read files from your device, or call other services. An app is a self-contained tool rather than an integration.
The reasoning behind that is in How a Traid assistant works.