Export the usage CSV, analyze burn with another model, then inventory routines one change at a time—pause don't delete—until a morning usage check gates nice-to-have work.
Sixteen named bots burned the week in days and you're deciding upgrade vs migrate.
Seven routines paused, all 16 bots kept, main bot checks the meter each morning before nice-to-haves.
Export your Grok Bot usage CSV.
Have another model (or a careful pass yourself) find which routines and cached context dominate burn.
Inventory bots and routines; change one thing at a time; pause don't delete.
Confirm each change before the next.
Add a morning usage check on the main bot so nice-to-haves run only when there's room.
What's good: CSV truth (here 91% cached context; four routines = 70% burn) plus pause-not-delete. The catch: the analysis loop used Claude paste-backs; you still act on the inventory.
Write-ups follow public posts and official docs. No likes, views, or dashboard numbers. Open the source to check the original. About