A lot of AI coding waste does not look wasteful while it is happening.
It feels like momentum.
Claude Code almost fixed the bug. Codex got close. Cursor produced a patch that looks plausible. So the natural move is to hit it again with one more prompt.
Sometimes that is the right call. Sometimes it is just an expensive loop with a friendly interface.
The small habit that helps me is checking usage before the retry, not after the session.
Before sending another AI coding prompt, I want to know:
- Is this still a small task, or did it quietly turn into a debugging session?
- Am I close to a usage limit or reset window?
- Did the last prompt improve the answer, or did it just produce a different guess?
- Would a narrower file, a failing test, or five minutes of reading the code make the next prompt cheaper?
The point is not to stop using AI coding tools. I use them constantly.
The point is to notice when the work has changed shape.
A retry prompt is cheap when the model has the right context and a clear failure. It gets expensive when the context keeps widening and the prompt is basically "try again."
That is why I like usage visibility in the menu bar. It puts the signal near the decision. If the meter is visible before the next prompt, it can actually change behavior.
I built TokenBar around that idea. It is a Mac menu bar app for keeping AI coding usage, limits, reset times, and spend visible while you work.
It is free to try, and TokenBar Pro is $15 lifetime:
The practical rule I keep coming back to:
If a retry would be your third similar AI prompt in a row, check the meter and tighten the task before sending it.
Top comments (0)