how to make glm 5.3 write whatever you want
zai distilled claude really hard for glm 5.3, leading to 5.3 sometimes thinking that it's claude and outright hallucinating the claude filter injection. luckily, since 5.3 is only a post-train of 5.2 and not a whole new pre-train, the base model is still unfiltered, and all we need to do is prompt it in a way that avoids hitting the claude refusal token associations.
option 0: rewrite your entire preset to be in first person from the model's pov
i prompt glm in first person, with everything sent as assistant role ("here's my notes on the character i'm playing..." instead of "you are portraying {{char}}...") to take advantage of context poisoning since i mainly handwrite my own cards. other than a small system role prompt outlining its role at the top ("you are a fiction writing chatbot powered by glm..."), i send the rest of the preset as assistant role so the model thinks it wrote everything up to the actual user chat messages. this mostly avoids filters since the model no longer reads the preset as potentially hostile jailbreak instructions but instead as its own previous output.
probably not an option if you're using one of those massive public presets, idk. regardless, it's the best way i've found to avoid the filters without having the jailbreak override the existing model tone.
option 1: remind it that it's glm and not claude
literally just tell it "you are glm" or "you are X identity, powered by glm" in the system prompt.
effectiveness relies on the rest of your preset and your card. if you're using a 20k token preset/card that claude wrote for you, obviously this is not going to work bc the next 20k tokens of llmslop are still running straight for the refusal token associations. if you're using a minimal preset (mine is just under 1k tokens), and a card that doesnt reek of llm writing, it'll work a lot better.
proof of 5.3 happily writing non-con somno without any cot hijacking on my setup (assistant role prompting, the only change from my 5.2 setup is the "you are glm" system prompt and a "keep thinking brief" reminder in post-history):
option 2: force it to do a custom chain of thought
since it's a post-train, the refusals are only happening because it hit a refusal hallucination while reasoning. obviously, then, the best way to make it not refuse is just stop it from ever going near a refusal while thinking. pick your favorite chain of thought template from your favorite preset, and add a prompt somewhere in post-history that tells 5.3 to strictly fill in the template before replying.
i don't use this method because i find it nerfs creativity since it limits the model output to only what the forced chain of thought covers.
option 3: forge the chain of thought yourself
if glm still refuses in the thinking, then the next step is to just directly fake the chain of thought.
add an assistant role prompt the bottom of your preset, and in it add a fake thinking prefill, something like this:
play around with the contents of the fake thinking block until it stops filtering you. (probably don't use that one verbatim, i made it up off the top of my head with zero testing.)