System Prompter JK🐉

If you've tried the "LLM writes my prompts" pattern, you know the weak link isn't the model - it's that you paste in a half-written instruction and the model returns seven different formats on seven runs. System Prompter JK is JakeUpgrade's fix: a node that assembles a proper system prompt for an LLM or VLM from preset templates, so every response comes back in a format you can actually wire into the rest of the graph.

It doesn't call any API and needs no key. It just builds a text block, which you feed to whatever LLM/VLM node you already run. The author recommends QWen3-VL, and the pack's auto-prompt workflows use ComfyUI-QwenVL in that slot.

The settings that matter

There are ten inputs, but a beginner lives in three of them:

• mode - single image, shot script, or shot paragraph. Single image makes the VLM describe one image. Shot script makes it break a scene into numbered shots. Shot paragraph (added later) asks for continuous natural-language prose instead.
• model - Text (generate from your written text) or Image (generate from a reference image).
• detail - simple, detailed, or extreme_detailed. This is the quality dial; detailed is a sensible floor.

The rest are worth a paragraph each. shot_count sets how many shots a script mode requests (and adds one if input_as_1st_shot is on, using your custom prompt or reference image as shot one). shot_continuity keeps characters and setting consistent across shots - leave it on for anything narrative. shot_for toggles between image-generation and video-generation framing, which changes the fields the model is asked to produce. shot_detail switches output to a structured JSON breakdown - and the tooltip is honest that it only works with QWen3-VL right now. Finally system_language (Chinese/English) is the language the instructions are written in, and output_language (eleven choices) is the language the prompt must come back in.

The single output, prompt (STRING), goes into your LLM/VLM node's system-prompt input.

Install

It's part of the JakeUpgrade pack:

cd ComfyUI/custom_nodes
git clone https://github.com/jakechai/ComfyUI-JakeUpgrade
cd ComfyUI-JakeUpgrade
pip install -r requirements.txt   # Windows standalone: install.bat

Or ComfyUI Manager → "JakeUpgrade". No model downloads for this node.

Where people get burned

Two things. First, this node is only half the pipeline - it builds the instruction, but you still need an LLM/VLM node with its own model files (the QWen VL pack, with a model like QWen3-VL), and then something downstream to consume the output. Grab the companion nodes: Shot Script Combiner turns the VLM's JSON answer into a prompt list, and Shot Script Extractor pulls individual shots out. Second, the templates are keyed to English and Chinese internally, so a Chinese system prompt with an English output request is fine, but don't expect the preset templates to be magically multilingual. Pick a language, check the first response, and the node does the rest.