Plan before you act

Intermediate · ~5 min · Prerequisites: None

What you'll do

Switch the assistant into Plan mode so it maps out exactly what it intends to do before touching anything. You review the proposal, ask questions, then flip to Agent mode to run it.

Flip to Plan mode

The composer has an Agent / Plan toggle next to the send button. Switch it to Plan before you send a message. You can also switch per-message mid-conversation — Plan and Agent turns can alternate in the same thread.

When you're in Plan mode, every message you send returns a read-only proposal: a numbered list of steps the assistant would take. Nothing in your workspace changes.

Review the proposal

Read through the proposed steps carefully. Common things to check:

• Scope — does it touch more than you expected?
• Order — are the steps sequenced correctly?
• Assumptions — does it state any preconditions you need to verify?

Ask follow-up questions or tell it to revise the plan: "Skip step 3" or "Add a step to notify me when it's done." The assistant revises the proposal without acting.

Approve and run

When the plan looks right, switch the toggle back to Agent and send a message like "Go ahead" or "Run it." The assistant executes exactly the steps it described. If it hits something unexpected it pauses and checks with you.

For multi-step actions, each mutation still triggers an approval card inside the run — you get a second checkpoint before each destructive or irreversible step.

When to use it

Plan mode is most useful when:

• You're about to do something that can't easily be undone (bulk edits, sending emails, creating many items).
• You want to understand the blast radius of a complex ask before committing.
• You're delegating a task to the assistant for the first time and want to validate its interpretation.

For quick, low-stakes requests — summarizing a doc, answering a question — Agent mode is faster and the extra step adds no value.

Try this next