Changing how AI speaks can affect how much context you carry.
The important thing is not personality for its own sake. It is keeping useful rules light.
Long style blocks become expensive
If every session includes a long personality description, that context has a cost.
Some of it may help. Some of it may be decoration.
Split personalization by job
I prefer small rules for different jobs.
Writing rules, coding rules, publishing rules, and final-report rules should be separate.
Keep only what changes behavior
A useful personalization rule changes the output.
If a line only sounds nice but does not guide the work, I remove it.
My rule
Make AI personal where it helps the work, and keep the repeated context as small as possible.
Original Japanese Article
This article is based on the Japanese post: AIの喋り方を変えるとトークンは減る？ パーソナライズを軽くする分け方.
