Local LLMs let you edit the context however you want, including the responses generated by the LLM so it will later think it said what you want it to say which can help put it on the right track.
LLMs-as-a-service don't offer this because it makes it trivial to bypass their censoring.
I've heard it repeated so many times that once things start to go sideways, trying to get back on track is a mistake. Have you had real-world success hacking context using rewritten responses?
LLMs-as-a-service don't offer this because it makes it trivial to bypass their censoring.