ZML_LLM 系统提示词

ZML_LLM_SystemPrompt is the smallest honest node in the ZML LLM set: one multiline text box, one string output, zero processing. You type a system prompt, it passes it through unchanged as a 系统提示词 (STRING). That's it. It exists so the personality and instructions for your chat have a clearly-labeled, reusable home in the graph instead of being welded inside a long user prompt.

If that sounds trivial, the default value tells you what the author actually uses this for - 你是一只猫娘 ("you are a catgirl"). The ZML pack comes from the Chinese anime-generation community, and a lot of these workflows are less "structured captioning pipeline" and more "chat with a character while you generate." Both are legitimate; the node doesn't care which you're doing. The system prompt is where you set the model's role, tone, output constraints, and - importantly - any "always return JSON / always answer in this format" instructions that you don't want to retype per generation.

Why does it need to be its own node at all? Because ZML_LLM_Chat takes system_prompt as a wired input (a STRING socket, not a widget). Having a dedicated node means you can: keep the system prompt separate from the user prompt (which changes every run), reuse the same role across several Chat nodes, and swap personalities by editing one box instead of digging through a long prompt. You could feed it from any text node - ZML's own text selection nodes, a text block from a saved image, whatever - but the dedicated node is the obvious starting point.

One genuinely useful interaction: the Chat node appends output-format requirements to your system prompt when a JSON schema is attached. So if you define a schema, your system prompt is the layer that gets the "please follow this format" directive injected on top. Keep the system prompt free of contradictory instructions and the JSON strategy stays reliable.

There are no failure modes worth warning about - if the output's empty, check the Chat node's wiring, not this one. It's the kind of node that only makes sense inside the ZML LLM stack, and it costs nothing to include.

Install the pack the usual way:

cd ComfyUI/custom_nodes
git clone https://github.com/zml-w/ComfyUI-ZML-Image

(Manager search "ComfyUI-ZML-Image" also works.) Restart, look under ZML 图像 → LLM, and note the default is in Chinese - change it or apply the author's English translation patch, whichever you prefer. No dependencies beyond the pack itself; this node literally never touches the network.