Skip to main content
Open the relevant project and write a follow-up in the composer. Be direct about the change you want:
  • “Use the research in this project and turn it into a one-page brief.”
  • “Keep the recommendation, but give me a version for an engineering audience.”
  • “Before making any external changes, show me the proposed actions.”
If an AI agent is busy, your message waits as Queued. Choose Steer now to redirect it immediately; steering does not use one of your daily cloud tasks. Type / in the composer for commands, or @ to mention a connected plugin so the AI agent uses it. Mentions guide the AI agent; chat permissions still apply.
  • /new starts a clean conversation.
  • /loop creates recurring work, and /goal keeps a project goal in focus.
  • /btw asks a side question without changing the task.
  • /plugins connects or manages apps, and /skills adds skills for your agent.
  • /settings opens agent connections, folders, and permissions.
  • /create asks for a polished result, /research for a cited investigation, and /summarize for a summary of the conversation so far.
For some requests, small controls appear above the composer so you can fill in details in place, such as dates, a time zone, a budget, a color palette, options to compare, or a Repeat schedule. You can also paste files and attach PDFs. The agent control sets permissions for the chat: Manual asks before actions, Plan inspects and plans without changing anything, Auto has a reviewer check actions before they run, so routine actions may proceed while risky or unclear ones ask you, and Full access acts without asking. Use Full access only when the task and workspace are safe for unattended changes. Open Activity to review what has happened, or Outputs to open a saved result. If an AI agent is waiting, open the decision card and either answer the question, approve the specific action, or decline it. Declining keeps the AI agent from taking that action; give it another direction if you want it to continue differently.