Stop digging the same hole
An AI coding session can reach a point where each new instruction seems to fix one thing and disturb another. Claude Code’s checkpoint feature offers a different move: return to an earlier point and try a clearer direction.
Run /rewind, or press Escape twice with an empty prompt, to open the menu. Select the earlier prompt, then decide what you want to restore. This is an existing capability introduced in September 2025, not a new model announcement.
Choose what should come back
Restore code keeps the conversation while reverting tracked edits. Restore conversation returns the discussion to the selected point while leaving current files in place. Restore code and conversation brings both back. Current documentation says code-restoration choices appear only when there are tracked changes to restore.
Think about a redesign that kept the right structure but sent the conversation down an unhelpful path. Keeping the code may be the useful choice. If the files themselves went wrong, keeping the discussion while restoring edits can preserve the reasoning you want to reuse. These are illustrative decisions, not results from a test project.
Make the next attempt easier to judge
After choosing a checkpoint, ask for one bounded change and name how you will check it. “Keep the layout; fix only the narrow-screen navigation” is easier to assess than another request to improve everything.
Checkpoint recovery has a boundary: shell-command changes and other edits outside its tracking are not a complete rollback target. Keep Git history for durable project recovery. Anthropic’s current documentation details additional limitations, including how subagent changes are handled.
Source published 2025-09-29. Coverage is based on the maker’s announcement and demonstration.