Why
After “no, not like that”, the context holds your question, the wrong answer and the objection. The model sees its previous answer as part of the conversation and patches it instead of solving the task afresh: same approach, different details. Three or four rounds, and the history has three wrong versions against one line of “not like that”.
Editing the original message throws away the wrong branch entirely: the model sees the question as if for the first time.
What it looks like
| Instead of | Do this |
|---|---|
| “no, I need PowerShell, not bash” | edit the question: “a PowerShell 5.1 script for Windows…” |
| “you forgot the port again” | add to the question: “port 8080 is taken, the service listens on 8081” |
| “too long” | in the question: “answer in 10 lines or fewer” |
Each such edit is something the prompt was missing. Keep the version of the question that worked: next time it works right away (A prompt, block by block).
Where to find it
In claude.ai and ChatGPT, the edit button under your message: the new answer goes into a separate branch, and the old
one is reachable with the arrows. In Claude Code, a double Esc with an empty input opens the rewind menu, where you
pick the message to go back to — with or without the code. In the API, send the history without the failed
question–answer pair.
When arguing helps: when the model is right and you aren’t. Asking “why this way?” is fine; repeating “no” without new information isn’t.