First requests

Composer basics

Type in the composer at the bottom of the panel.

Example prompts

These two prompts show the range of what a first message can ask for: one that only reads the codebase, and one that changes it.

A read-only prompt, useful for orienting in an unfamiliar repo:

Explain how authentication is wired in this repository. Do not modify files.

The explicit "do not modify files" isn't strictly required, since a read-only question doesn't trigger edit tools on its own, but it removes any ambiguity about intent and makes the prompt safe to fire off without a second thought.

A prompt that expects edits and verification:

Fix the registration validation bug and run the relevant tests.

This one asks for both a change and proof the change works: the agent edits the file, then runs the test suite as verification before considering the task done.

The first is safe to run in Plan mode since it makes no edits. The second needs Build mode, since it edits files and runs commands.

What happens next

• Assistant text streams in, along with a separate stream of tool activity (file reads, edits, commands) attached to the same message. See conversations.
• Any edit or command that needs your say-so shows a permission prompt before it happens.
• A context meter tracks how much of the model's context window the conversation is using.

If you want more control before the agent starts editing, switch to Plan mode first and ask it to investigate; then switch to Build mode once you're ready for it to act.