Why don’t more people adjust the custom instructions for model behavior?

Custom instructions can be a bit of a band-aid solution. Even if you tell the model not to use a specific word, it'll often slip back into those habits. It's not as reliable as we'd like for both simple and complex instructions.

It's possible that users have tried adjusting the instructions multiple times and still see the same problematic behavior. Custom commands can help, but they’re not a one-size-fits-all fix—the model doesn't always follow them consistently.

Honestly, my gut reaction is to think that the users are part of the problem. It reminds me of troubleshooting tech issues back in the day—sometimes it's just about knowing the right way to ask for what you need.

I think a lot of it comes down to style preferences. If the custom instructions were more like some of the other models, where I can set my own templates or presets, it would be way easier to work with. As is, trying to fix the model's default style with just global instructions can lead to some unintended results.

A lot of users seem to forget that the system prompt actually takes precedence over custom instructions. You might set it up to behave differently, but it'll just revert to its original training over time. It's kind of annoying since you expect it to listen but then it just goes back to its old habits.