Undo a bad AI answer instead of correcting it

The AI gives you something wrong. Maybe your ask was worded badly, maybe it confused the AI, maybe the answer is just bad or maybe you've changed your mind since you sent it. The natural move is the one you'd make with a person. You tell it what's wrong and ask it to fix it.

That costs more than it seems. The wrong answer doesn't leave when you correct it. It stays in the conversation, and the AI rereads the whole conversation every time it answers, so your fix gets built on top of the attempt you just rejected, with the misunderstanding that caused it still sitting there. A conversation that has taken three corrections is carrying three wrong answers and the ask that confused it in the first place.

The answer is to treat it as an undo, so the AI sees a conversation where the wrong turn never happened. Thariq Shihipar, who works on Claude Code at Anthropic, made the case in April 2026 in "Using Claude Code: session management and 1M context", a guide to keeping long working sessions healthy. His example is an approach that failed after the AI had read the files it needed. "Rewind is often the better approach to correction," he wrote, and the same logic holds for every other way a turn goes wrong.

Correcting keeps the mistake in the conversation

An AI doesn't remember a conversation the way you do. Each time it answers, it reads everything said so far and works from all of it. That includes your messages, its own answers and anything it looked at along the way. That's the context window, the text the AI is working from at that moment.

So a correction adds to the pile rather than replacing anything. The bad answer is still there, and so is your first ask, worded the way that confused it. The AI now has two versions of what you want and one example of getting it wrong, and it works from what's already on the page. Watch for a next attempt that looks like the last one with the complaint patched in.

Rewind is an undo for the conversation

In Claude Code, the AI coding tool that runs in your terminal, type /rewind, or press Escape twice with nothing typed, then pick the message to go back to in the rewind menu. The list shows only your own messages, each one the start of a turn, meaning one ask and everything the AI did to answer it. The menu then asks what to do there and offers up to six options. Three restore something, the code and the conversation, the conversation alone or the code alone. Two summarize, either from the chosen message onward or everything before it, and the last is Never mind.

Pick the point to go back to with care. If the AI read files or looked things up in an earlier turn, go back to the message after that turn, so it keeps what it read and loses only the bad attempt. When the reading and the bad attempt happened in the same turn, they go together.

Pick the conversation alone when the files are fine and only the talk went wrong. Pick code and conversation when the bad turn also edited files, so those edits come back out too. The two code options only appear when the AI edited files with its own edit tool after that point. Summarize from here is for a turn that went wrong but found something worth keeping, because it condenses the messages from that point into a short summary instead of dropping them. Summarize up to here condenses everything before a point to free up room, which is a different job from undoing.

After any option that restores the conversation, and after Summarize from here, the prompt you sent from that message comes back into the input box, so you can reword it there. Say the thing you'd have said in the correction, but say it up front, as part of what you want, so the AI never sees the wrong version.

In a chat app, check whether you can edit an earlier message and send it again. Where you can, it does the same job, because the answers after it drop out of the conversation.

Rewind throws away the whole turn, so it isn't always the right move. When the answer is mostly right and one detail is off, a one-line correction is cheaper than a do-over, and the good part stays. The line to draw is whether you'd want the AI to keep reading the answer. If you would, correct it. If you wouldn't, rewind.

What you get and what you don't

You get a conversation that holds only what you meant, so every answer after the rewind is built on the ask you wish you'd sent the first time.

You don't get back what the AI did outside its own file edits. Rewinding doesn't undo changes made by commands it ran, such as a file it deleted or moved from the command line, and edits made by helper agents it ran in the background usually stay too. So save your work to version control, such as a git commit, before a turn that might go wrong. And you don't get an AI that learns from the mistake. The rewound turn is gone, lesson and all. If the lesson matters beyond this conversation, write it into the rules file your agent reads first, the way Write AGENTS.md rules that survive the session describes.

The next time an answer comes back that you'd rather the AI never saw, don't reply to it. Rewind.