Irrelevant or unhelpful context increases the chances of mistakes. Let me explain how to handle this with the rewind feature.

If Claude makes a mistake, you're better off rewinding and resuming the conversation before it made the mistake. This is a little bit unusual compared to working with a human being. If a human developer makes a website and puts it in blue, you say "No, it should be in red," they go back and change it to red, and they don't spontaneously remember that they did it in blue before.

But Claude has this massive context with both "blue" and "red" in it. The red is just later in the context. The probability of remembering red versus remembering blue has diminished, and you're starting to get a higher probability that it might spontaneously start doing blue again.

So we need to use the rewind feature. Every time Claude is about to start a new prompt, it checkpoints the state of any files it's about to modify. Between each step in your conversation, you can rewind back to the state at that particular point. You can rewind both the conversation and the code back to that state.

When you rewind, it's as if the later instructions never happened. There's no fast forward, so you'll need to redo anything you did after that point. But the key advantage is that Claude's context is now clean. It never saw the confusing or incorrect instructions, so it won't get confused about what you want.