When I am using Claude Code, Codex, or Cursor, the expensive mistake is usually not one giant prompt.
It is the fifth small task after I stopped paying attention.
A bug fix turns into a refactor. The refactor turns into a test cleanup. Then I ask for one more pass because it feels cheap in the moment.
By the time the session feels heavy, the usage is already gone.
So I started doing a tiny pre-flight check before opening another task with an AI coding agent.
My quick check
Before I ask the model to do more work, I look for three things:
- How long has this session been running?
- Did the last prompt move the code forward, or just move text around?
- Am I asking for a clear next step, or am I asking the model to keep exploring?
That third one matters the most.
A clear next step sounds like:
Add validation for this settings field and update the failing test.
A drifting prompt sounds like:
Look through the project and improve anything weird.
The second one can be useful, but it is also where token usage quietly explodes.
Why usage visibility changes behavior
I do not want a huge dashboard while coding. I mostly want one visible signal that makes me pause before I burn another chunk of context.
If usage is hidden in a billing page or a web console, I will not check it.
If it is sitting in the Mac menu bar, I actually notice it.
That small bit of friction helps me decide whether to:
- keep going
- tighten the prompt
- start a fresh session
- stop and read the code myself
The goal is not to avoid AI coding tools. I use them constantly.
The goal is to stop treating every prompt like it has zero marginal cost.
The habit that helped
My rule now is simple:
Before starting a new AI coding task, I check whether the previous task earned the usage it spent.
If it did, keep going.
If it did not, write a smaller prompt or do the next step manually.
I built TokenBar around this kind of habit: live LLM token and usage visibility in the Mac menu bar for Claude Code, Codex, Cursor-style workflows, and other AI coding sessions.
It is free to try, and Pro is $15 lifetime.
If you use AI coding agents every day, the cheapest token is usually the one you decide not to spend.
Top comments (0)