ThinkingLLM System Prompt Preset

One input, one output, no model involved, no VRAM, no API key. It's the most boring-looking node in the pack and it's quietly one of the more useful ones, because it hands you the actual text behind ThinkingLLM's preset dropdowns.

Here's the thing: the presets are the real product. The "👁️ Remove (Mask Area)" preset isn't a label - it's a long, carefully written instruction that tells the model to describe what should fill a masked region as a natural background continuation. The MiniMax H3 and LTX 2.3 video presets encode the official prompt-writing structures those models expect. All of that lives in the pack's AILab_System_Prompts.json, and this node just reads the file and spits the selected preset out as a plain STRING. That's the whole mechanism, which is exactly why there's nothing to go wrong.

Why you'd reach for it

The main ThinkingLLM nodes have this preset selection baked in, so why would you need the string on its own? Because it plugs the same instructions into other things:

• Feed it into the ThinkingLLM API (OpenAI Compatible) node's system_prompt input, and your cloud model follows the exact same instructions as your local run. Two models speaking one dialect.
• Wire it into any LLM node you're already using that accepts a system-prompt string.
• Connect it to a Show Text node and actually read what the presets ask the model to do. If you've ever wondered why a video prompt preset produces such structured output, reading one is the fastest education you'll get.

LLM-assisted prompting is a twentyfold-growth corner of ComfyUI for good reason: the text encoder of every modern model is itself an LLM, so having an LLM write the instruction is translation between two things that speak the same language. This node is the low-effort way to make sure the translation is consistent.

The inputs and output

Just one input: preset_prompt, a dropdown with a couple dozen presets. Among them: the image-only and custom-prompt options, the mask-removal preset, the LTX 2.3 and MiniMax H3 video scene/reference presets, Wan 2.2 NSFW timeline and scene presets, and the pure-description ones (Tags, Simple/Detailed/Ultra Detailed Description, Cinematic Description, Video Summary).

Output: system_prompt, a single STRING with the full text for your selection. It's not the model's output - nothing generates anything. Think of it as a constant you can route anywhere a system prompt is expected.

Install and gotchas

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/goodguy1963/ComfyUI-ThinkingLLM.git
cd ComfyUI-ThinkingLLM
pip install -r requirements.txt

or ComfyUI Manager → search ThinkingLLM, then restart. (Yes, this utility still pulls the pack's heavy requirements - there's no way to install just one node.)

Because it reads the JSON at runtime, updating the pack changes what a given preset emits. Old workflows don't break, they just get the new instructions - usually an improvement, occasionally a surprise if you update mid-project. If the node ever errors, it's the preset file failing to load, which after an update means restart ComfyUI once. And no, there's no thinking toggle or token streaming here - that's the charm. It's a string reader. Use it like one.