Do long system prompts make models lose track of rules?

It is so easy to fall into the trap of thinking that adding more text fixes bad outputs. We used to write massive instructions packed with edge cases and the model would end up ignoring half of them anyway.

Trimming everything down to clear bullet points and simple variables made the responses way more reliable. Cutting out the fluff really does force the model to focus on the actual task.

Do you prefer keeping your system instructions short or do you like adding strict guidelines for formatting?