Hi, I’ve found it kind of funny to watch Meta push soul.md files into the spotlight and see people's reactions to them.

I tried making some interesting "souls" for Muse.

Even though Meta is pushing Muse as this unique ever-evolving personal assistant, it still seems to have an iron grip on its personality.

I have never had such difficulty getting a model to use a given persona or even follow basic preferences like writing style.

Muse does have a long system prompt defining its disposition, but this is also true of other assistants like Claude. (Claude is quite good at following soul.mds IME).

So I am not sure if it is the system prompt, the training itself, how the context is processed, or something else.

Maybe the model itself is just not as good at following custom instructions?

Given this, YMMV.