Load gpt-oss Prompt Designer

The name is doing two jobs, one of them a lie. "Load gpt-oss Prompt Designer" doesn't load anything - no model file, no checkpoint, no weights. What it does is talk to a local Ollama instance running OpenAI's open-weight gpt-oss:20b model, hand it a topic, and get back a finished Stable Diffusion prompt as a text string. It's a one-node prompt generator, not a loader, and it needs no API key because nothing ever leaves your machine. If you're tired of staring at an empty prompt box, this is a genuinely low-effort way to outsource the staring.

This lands right in the middle of a trend the community has fully adopted: having an LLM write your prompt instead of writing it yourself. It works because the encoder is reading your prompt as an instruction, so an LLM writing that instruction is just translation between two things that speak the same language. Dedicated local prompt-writing node packs now ship per model family and get real attention on release. This one is from a single Japanese author, has a single commit, and essentially zero community footprint - it's a small, honest tool, not a platform.

How it works

Under the hood it's embarrassingly simple. The node calls the OpenAI Python client pointed at http://localhost:11434/v1 - that's the local Ollama API endpoint, with a dummy api_key value that Ollama ignores. It sends your system prompt and user prompt as a chat request for model gpt-oss:20b, and returns the model's text output as a STRING. No external API, no cloud round-trip, nothing you haven't already run locally. The one wrinkle: the model name gpt-oss:20b is hardcoded in the source, so you can't switch models from the UI without editing nodes/nodes.py.

The inputs

Only two, and both are plain text:

system - the role the model plays. Default: You are Stable Diffusion Prompt Designer. You can leave it.
user - the actual request. The default is a Japanese template asking for an English prompt on a theme (~30 words, ending with ## Output). Swap the theme line for whatever you want to generate.

Both raise a ValueError (in Japanese) if you run the node with them empty, so don't clear them.

Wiring it in

The single STRING output goes straight into a CLIPTextEncode positive prompt (or any text input you like). Run the node, wait for the local model to think, and the generated prompt lands in the text box. Note the node is marked as an output node, so it also prints the result to the console.

Install

The README targets Google Colab, but on a desktop the path is the same shape. First the node:

cd ComfyUI/custom_nodes
git clone https://github.com/educator-art/ComfyUI-gpt-oss-PromptDesigner

Or search "gpt-oss" (or "PromptDesigner") in ComfyUI Manager. Then Ollama itself:

curl https://ollama.ai/install.sh | sh
ollama pull gpt-oss:20b

Keep ollama serve running in the background before you hit the node. One gotcha the README never mentions: there's no requirements.txt in the repo - it assumes Colab, which ships the openai package. On a desktop install you'll likely hit ModuleNotFoundError: No module named 'openai' and need to pip install openai into the same environment ComfyUI runs in.

The catches

The big one is VRAM. gpt-oss:20b is a ~14 GB model and the author targets an L4 with High Memory for a reason - an 8 GB card will not run it. The pull downloads 12–14 GB depending on the quant, and the first generation is slow while the model loads into memory. If it fails with a connection error at localhost:11434, Ollama isn't running; if it complains the model is missing, you skipped the pull.

Worth saying plainly: for a custom node, this one is unusually safe. It's GPL-3.0, it only calls your local Ollama, and there's no exfiltration surface to speak of - a nice contrast in a registry with a documented history of malicious LLM nodes. It won't change your life, but for "I can't think of a prompt," it's the one you can trust to not phone home.