Your prompt is what you type. The model reads it along with its context, which is the conversation so far plus any files it has read.

When you attach a file with @, it’s in the context from the start, so the agent doesn’t have to guess which file you mean.

A good prompt covers four things.

Part	Example
Goal	write actions.md
Context	from @notes.md
Limits	don’t change notes.md
Done when	one line per action item, with who and when

Try it

1. The project has three files. Open notes.md in the Editor tab. It’s a messy meeting note.
2. Start with a vague prompt, clean this up. Which file did the agent pick, and what did it do with it?
3. Press Reset lab and ask again, this time with the goal, context, limits and “done when” from the table. The prompt is already in the Try box.
4. Compare the two results. Which one did you actually want? The agent knows only what you tell it.

The prompt line accepts four kinds of input, and the first character decides where each one goes.

Type	What it does	Use it when
/help	lists commands and keys	you forget one
/context	shows what fills the model’s memory	replies get worse
/compact	folds the conversation into a summary	the context is nearly full
/clear	forgets the conversation but keeps the files	you start a new task
/diff	shows every change since the lab started	before you move on
/init, /memory	write or open AGENTS.md	you want the agent to remember the project
/model, /reset	show the model; start the lab again
!ls	runs a shell command without the agent	you want to look for yourself
shift+tab	default → accept edits → plan
esc	stops the agent	it’s heading the wrong way

Try it

1. Type /help. If you type just /, a list appears and narrows as you type.
2. Run !ls to see the files left from the last lab.
3. /context shows what’s in the context, part by part. Besides the conversation, you’ll see the tool’s own instructions.
4. Finish with /model.

Here’s a small task that uses what you just learned.

Try it

1. Press Type it. The prompt appears on the line, and the status line under the terminal switches to plan mode. Press Enter. In plan mode the agent can’t change files, so it answers with a plan.
2. Under the plan it asks Would you like to proceed? Choose Yes, and manually approve edits so you see the change before it’s made.
3. Read the diff for packing.md, then choose Yes.
4. /diff shows every change since the lab started, and /context shows what all this cost.
5. Type /clear, then ask What's in packing.md? The conversation is gone but the file is still there, so the agent reads it again.