Escaping "</think>" and other sensitive tokens from the input?

[l29ah]

I would like to be able to discuss every stuff imaginable, including "</think>", with a LLM, but inclusion of </think> in the prompt confuses it (today it is GLM-4.5-Air), apparently due to the fact that the tokenizer uses the same token as its "thinking" stuff it was trained to interpret in a special way. How can i ask the /tokenize to use other tokens for that? parse_special option has no effect on it.

[CISC]

You would have to tokenize it in two or more parts and use those tokens, as parse_special or not, the </think> token is the shortest possible token representation.