Many LLMs can interpret invisible Unicode Tag characters as instructions and follow them (eg invisible comment or text in a GitHub issue).
I wrote about this a few times, here a recent example with Google Jules: https://embracethered.com/blog/posts/2025/google-jules-invis...