Rewind in action — and real YOLO mode
Make checkpoints concrete with a live rewind, then meet --dangerously-skip-permissions: the mode that asks for nothing and started a whole movement. How it differs from accept-edits, and how to run it without regret.
Time to make the last page concrete. We'll drive a real rewind, then meet the mode that started a whole movement: YOLO — running Claude with permissions switched off entirely.
01Rewind, for real
Set the stage: summarise the project, ask for a code review written to docs/review.md,
and watch it work with Ctrl+O for the detailed trace. In the demo the review made a
familiar mistake — flagging an "exposed API key" that was actually safely in .gitignore.
Rather than correcting it outright, we nudged: "Are you sure the API key is exposed?" — and it worked
the problem out for itself and fixed the files.
Now we can rewind to before that hint and see the difference.
Run /rewind
A list of steps appears, much like resuming a session. The arrow marks "current."
Press ↑ to the step you want
Move back past "Are you sure…?" to "Please do a code review," or further to "Summarise the project."
Choose what to restore
Both conversation & code · just the conversation · just the code · or cancel. Pick both.
review.md to what it said before the hint — the "exposed key" is back.Check review.md afterwards: the file has been reverted alongside the conversation.
Both moved back in time together — exactly the checkpoint behaviour from the last page, now visible.
02What YOLO is not
When we pressed Shift+Tab for "accept edits on," it was tempting to call that YOLO. It isn't. Accept-edits merely auto-approves file diffs instead of stopping at each one. Real YOLO goes much further.
| Accept-edits (Shift+Tab) | YOLO (bypass permissions) | |
|---|---|---|
| Approves file edits | Yes, automatically | Yes |
| Approves running commands | No — still asks | Yes — asks for nothing |
| How you enter it | A keystroke, mid-session | A launch flag, deliberately |
| Risk level | Low | Real, though low with a frontier model |
03Real YOLO mode — bypass permissions
Commit first — always — then launch Claude in a mode where it asks for nothing:
git add . && git commit -m "before YOLO" claude --dangerously-skip-permissions
The flag is verbose on purpose — Anthropic really wants you to know you're doing something risky. You'll see a red warning that Claude won't ask before running potentially dangerous commands, and that this is best used in a sandboxed container. Proceed and you take responsibility.
With a frontier model like Opus the chance of something destructive is very low — there are no
well-known stories of it running rm -rf on someone's drive. But it's a statistical
model, so the probability isn't zero. Do it only if you understand and accept that. Proper sandboxing (next
week) is how the pros make this safe.
Please improve the UI of this project — make the horizontal layout look better, use icons instead of delete buttons, and use the horizontal space properly. Make your changes, test everything, and let me know when done.
That's a nicely contained instruction, which is why UI work is a sensible first YOLO. In the demo it ran for 7 minutes 44 seconds — unattended — and came back with a genuinely sharper, mobile-responsive interface: cards flowed into two neat columns, delete actions became crisp little bin icons, and the AI assistant still worked. People who do this for a living give far more grandiose instructions and walk away for an hour.
The moment YOLO finishes and you're happy, snapshot it: git add . && git commit -m "after UI revamp".
That single habit is what makes fearless building safe.
✓ Key takeaways
/rewindsteps back through prompts; you choose to restore conversation, code, both, or neither.- A rewind moves the chat and the files it changed back together — but not files touched by scripts.
- Accept-edits (Shift+Tab) only auto-approves edits; YOLO auto-approves everything.
- Enter YOLO with
claude --dangerously-skip-permissions— always commit first. - Risk is low with a frontier model but non-zero; contained tasks (like UI) are the place to start.
- Commit again the instant you're happy. Never trust uncommitted work.