Skip to content

System prompt

Before you type a word, the agent has already been given its instructions: the system prompt. This screen shows you those instructions, word for word, and lets you add your own. In the screen’s words: “What the main agent is told before every conversation. The base prompt is the same for every session; your additions layer on top of it.”

Control → System prompt in the sidebar.

“Exactly what the agent sees, read-only — the same honesty as the context panel.” You can read it, but you can’t edit it here. (The context panel is where a session shows everything the model can see right now: see Your first session.)

The first time you open this screen, the box may be empty: “Run a session once and the resolved prompt is captured and shown here (it stays visible afterwards).” That’s because the full prompt is only put together when a session actually runs. Once one has, the app keeps a copy, so the box stays filled after a restart. A copy kept from an earlier session is marked “from your last session”.

“Appended to the system prompt for every workspace.” Type your instructions in the box and click Save. “Saved.” confirms it.

Your additions reach new or restarted sessions, not the one already running. For instructions that belong to one project only, the screen points you to that project’s AGENTS.md instead: a plain text file in the project folder, read whenever a session starts there.

“Let the agent see which files you have open, and which terminals it has running.”

Include open files and terminals is on by default. The switch reads On while it’s on. Click it to turn it off. It tells the agent which files you have open, as a list of paths: “never their contents.” That’s what lets you say “this file” without spelling out the path. It’s sent only when the list changes, and it takes effect on your next message.

It also lists the terminals the agent started itself, so it can find a dev server it opened earlier in a long conversation, and the address of the browser pane it opened, never the page’s content. If you turn this off, the agent loses that too: “the terminal keeps running and you keep seeing it, but the agent may lose track of it.”

  1. Open Control → System prompt.
  2. Type your instructions under Your additions.
  3. Click Save.
  4. Start a new session, or restart one, to use them.
  1. Open Control → System prompt.
  2. Under Open files and terminals, click the switch so it reads Off.
  3. From your next message on, the agent no longer gets the list of your open files, its terminals or its browser pane.

Every new session is given the base prompt with your additions on top. With Include open files and terminals on, each message you send also carries the list of files you have open, the agent’s terminals and its browser pane, whenever they’ve changed.

  • Memory: notes the agent keeps about you and your projects.
  • Permissions: what the agent may do without asking.
  • Terminal: the terminals the agent can start.
  • Models: the model that reads this prompt.