The big problem examples have are “contamination”
The model learns shape, but also can mistakenly think a user said something they did not.
Ideally carefully crafting system messages can do the trick , that would be my first resort
Examples in a system message can lead to less leakage cause it can be clearer to a model it is just an example
A minimal thing I would recommend Don, is writing your system message in Hungarian, it could help
Maybe even try giving an xml tool example or two in the system message?