IMO a tinyest language model can be ran on messages over a certain length to determine if it’s Nicole spam or not.

I'd strip out everything other than the letters, then check maybe a dozen regexes for different parts of Nicole's typical message. If the number that match is over a certain threshold I'd flag it as spam. That way obfuscation with spaces or special characters won't work (as long as they keep the same letters) and small variations in the message will still be flagged.