Skip to main content
The agent does its best work with clear, focused messages. These tips help you get there in fewer turns.

​
Start small, then iterate

On a new project, the agent builds the smallest working version of your idea first, on purpose: usually one screen with sample data and little polish. That way you see something running quickly. From there, add one piece at a time. “Add a settings screen with a dark mode switch” works better than listing ten features in your first message.

​
One clear change per message

  • Ask for one change at a time. Smaller requests are easier to check and easier to undo.
  • Say what you see and what you expected. “Tapping Save does nothing. I expected it to go back to the list with the new item added.”
  • Attach a screenshot when something looks wrong. See Screenshots, files and voice.
  • Name things the way the app does. Use the screen titles and button labels from your app.

​
Agree on an approach first

For bigger changes, switch to Ask or Plan before you use Build. The agent reads your project and explains how it would do the work. In Remote chats, which run in the cloud, Plan and Ask can’t change files. When the plan looks right, switch to Build and ask it to build the plan. See Build, Plan, Ask and Design.

​
Diagnose first, then fix

When a bug survives a couple of attempts, ask the agent to find the cause before it changes more code. Use Ask mode, or say so in your message:
When the explanation makes sense, switch to Build and ask for the fix.

​
Accounts and shared data

By default, apps keep their data on the device. The agent doesn’t add sign-in, accounts or a server unless you ask for them. If your idea leaves this open, for example an app where people might share lists, the agent asks you one plain question first, such as whether other people should be able to sign in and see the same data. Answer it and the agent builds accordingly. To skip the question, say it up front: “Each user signs in and sees only their own notes.” See Data, accounts and Newly Backend.

​
Automatic fixes

When your app fails to build, hits an error or crashes in the preview, the agent starts fixing it without you clicking anything. You’ll see a message such as “The app build hit a problem. I’m fixing it and retrying automatically.”
  • Automatic fix turns use credits, like any other agent work. See How credits work.
  • To prevent a repair, press Stop. Automatic fixes stay off until you send your next message.
  • The agent doesn’t retry forever. In Remote chats, after the same build failure happens 3 times in a row, it stops and explains the problem instead.
  • Some problems need you, such as signing in or connecting an account. The agent doesn’t try to repair those on its own.

​
Standing instructions for your project

To give the agent instructions it follows in every chat, add a file named AGENTS.md or CLAUDE.md at the top level of your project. Each file can be up to 20,000 characters. Good things to put there:
  • Code style and naming you want kept
  • How the app is organized
  • Libraries to use, or to avoid
These files guide how the agent writes code. They don’t change how the agent builds or runs your app. To add one, ask the agent: “Create an AGENTS.md that says we always use metric units and never add new libraries without asking.”

​
You don’t need to pick skills

The agent brings in the expertise a task needs on its own, such as Expo and React Native know-how or App Store guidance. You don’t need to choose or turn anything on.

​
Going back is cheaper than arguing

If a change went in the wrong direction, don’t spend several messages talking the agent back out of it. Use Restore to here under an earlier message to put your code back, then try a clearer message. See Undo and Restore to here.