You have explained it three times. It keeps almost doing the right thing. You are about to explain it a fourth time.
Don’t.
Why correcting stops working
Everything in the session is still there: the misunderstanding, the wrong attempt, your correction, its apology, the second wrong attempt. That whole pile is the context the next answer is built from.
You are not talking to a fresh mind that has now understood. You are talking to a conversation that has spent twenty turns exploring a wrong idea in detail. The wrong idea has the loudest voice in the room.
The three-strike rule
Same problem, three corrections, no convergence — stop. Not because it cannot be fixed, but because the expected cost of the fourth attempt is now higher than the cost of starting clean.
What starting over actually means
Throw away the code. Not “keep the good bits” — that is how the confusion survives. If the work is committed, git restore or a checkout of the last good commit puts you back in a known state in one command. Claude Code also keeps checkpoints of the session, and /rewind lets you step back to an earlier point, which is handy when the mess is recent and small.
Throw away the conversation. Start a new session, or clear the current one. A fresh context is the point of the exercise.
Keep what you learned. This is the part people skip. Those three failed attempts told you something: which file it kept going to, which assumption it kept making, which term you used that meant something different to it.
Rewrite the request with what you know
The second attempt should not be the same request in a new window. It should be the request you now know you meant.
Earlier attempt: Make the cart update when the discount changes.
Rewritten: In
src/cart.js,applyDiscountshould recalculate the total and return a new cart object. Do not touch the render layer — the UI already re-renders when the cart object changes.
The second one closes the door the first one left open. You only knew which door to close because you watched it walk through the wrong one three times.
Quick check
What is worth keeping from a session you are abandoning?
The signals to watch for
Any of these means stop and reset:
- It reintroduces something you explicitly ruled out.
- Its explanations get longer while the change gets smaller.
- You are editing its output by hand more than you are asking for changes.
- You catch yourself typing “no, as I said before…”.
- The change has grown to files you never mentioned.
Only cheap if you prepared for it
Every word above assumes a committed, known-good state behind you. Without it, “start over” means “reconstruct my project from memory”, and that is a different and much worse afternoon. Commit before you start, and after every turn you liked.
Say it out loud
Abandoning a session feels like failure, which is why people keep pushing. It is not. It is the cheapest correction available once the conversation has gone wrong, and the people who are fast with these tools do it without hesitating.
Can I paste the good part of the old conversation into the new session?
What if it fails the same way in the fresh session too?
Try it yourself
The third correction on the same problem. Decide what happens next.