Skip to content

Choosing a Work Mode

Mode Behavior Use it when
Agent You say it, it does. Reads, writes, runs commands, delivers You know exactly what you want
Plan Proposes an approach, waits for your approval, then acts Complex or irreversible work
Ask Talks only. Touches nothing Exploration and advice

In Agent mode it runs to completion: edits files, executes commands, installs dependencies — without checking in.

Good for:

  • Small, well-defined tasks (“add unit tests to this function”)
  • Work you’ve already thought through
  • Anything reversible (you have git)

Bad for:

  • Deleting, overwriting, or publishing
  • Work where you haven’t decided on the approach

⚠️ Hard-won lesson: told to “clean up unused files,” Agent mode really deletes things. For destructive operations use Plan mode — or first have it list what it plans to remove.

Plan mode outputs an execution plan — steps, files touched, trade-offs — and waits for approval.

Always use it for:

  • Refactors and architecture changes
  • Edits spanning many files
  • Database migrations
  • Anything you can’t undo afterward

The cost is one extra round trip. Compare that to restoring deleted files.

Ask mode is read-only, which makes it the safest way to explore.

Good for:

  • “How is auth organized in this project?” — it reads and explains
  • “Which of these two options fits my situation?”
  • Clarifying requirements before you commit

People open Agent mode just to ask, and the AI “completes” the task by editing a few files. Asking is what Ask is for.

Does this require changes?
├─ No → Ask
└─ Yes ↓
Can you undo it?
├─ Yes, and the approach is settled → Agent
└─ No / approach undecided → Plan

For non-trivial projects, mixing is optimal:

  1. Ask: understand current state
  2. Plan: confirm approach and blast radius
  3. Agent: execute now that the plan is settled

Nothing gets changed behind your back, and you’re not stuck approving trivia.

Your First Task.