“One more prompt” is the vibe-coding version of one last commit. The feature is nearly working. One more request to the model should finish it. Then the fix breaks something else, and the next prompt fixes that, and an hour goes by.
The loop is normal, and most of the time each prompt does move things forward. It turns unproductive in a recognizable way: the model keeps proposing the same fix, undoes its own earlier change, or reintroduces a bug you already solved. That usually means it is missing a piece of information, or the conversation has grown so long that early details are fading out of its context window.
Ways out that work. Stop and go back to your last working commit rather than piling fixes on a broken state. Start a fresh conversation and describe the problem from scratch in a few sentences, including what you already tried. Paste the exact error message. Ask the model to explain what it thinks is going wrong before it changes anything. Break the task into a smaller step you can test on its own.
Sometimes the best prompt is to read the code yourself for five minutes. You do not need to understand all of it to spot that a function is called twice, or that a file the error mentions does not exist.
Engineers have the same loop with commits and the same fix: step back, narrow it down, try again from somewhere known to work.
In the wild
Third time it’s brought back the same bug. Reverting to this morning’s commit and starting a clean session with the error message.
