I started treating AI context like inventory, not storage

One workflow change that helped me with AI coding was simple: I stopped treating the context window like free storage.

It is more like inventory.

Every stack trace, pasted file, old TODO, failed approach, and half useful answer that stays in the thread has a carrying cost. It may not break the session immediately, but it changes the next answer. It makes the model sift through more history. It makes retries feel normal. It makes the session look productive while the useful signal per token keeps falling.

The trap is that bigger context windows make this easier to ignore.

A larger window is great when the extra context is actually relevant. It is not great when it lets dead context survive for another hour.

The practical habit I am trying to build is a quick context check before the next prompt:

1. Is the next question still about the same task?
2. Did the model already solve the hard part?
3. Am I keeping logs or files around because they are useful, or because I am too lazy to trim them?
4. Would a fresh session with a 10 line brief be clearer than dragging this thread forward?
5. Am I using a bigger model because the task needs it, or because the context is messy?

This is also why I care about live token counting more than after-the-fact billing dashboards.

A dashboard tells you what happened.

Live token visibility changes the moment you decide whether to keep going, trim context, restart, or switch models.

That is the small reason I built TokenBar. It is a macOS menu bar token counter for AI work. Not a full analytics suite. Just a visible meter while you are in the session, so context bloat is harder to miss.

If you use AI heavily as a founder or developer, I think "what am I carrying forward?" is one of the highest leverage questions to ask before the next prompt.

TokenBar: https://tokenbar.site/