⭐ Star Ollama Sysprompter (JC)

The local-LLM prompt-expansion workflow is a classic: Ollama runs a small model on your GPU, you feed it "describe a cyberpunk street scene with rain," and it hands back a rich paragraph that goes into your image model's text encoder. It's genuinely good - and it's also where every inconsistency lives, because each run you're hand-writing a slightly different system prompt and the LLM drifts between "elaborate prose" and "comma-separated tag soup."

StarOllamaSysprompterJC is a string factory that fixes the drift. It builds the system prompt for you - consistently, every time - and it also produces a style-specific "detail prompt" you can append to your actual generation prompt. It doesn't run Ollama, and that's the point: it writes the words you feed into your Ollama node, so the LLM behaves predictably.

What it outputs

Three strings:

• system_prompt - the instructions for the LLM: how to expand prompts, what the output should look like. This is your chat node's system message.
• detail_prompt - a style-flavored elaboration that rides along with your generation prompt, so the image model gets the same style description the LLM was told to aim for.
• style_name - the selected style as plain text, handy for metadata or a filename.

What you set

• style - 26 presets: Photorealistic, Watercolor, Oil Painting, Cyberpunk, Film Noir, Pencil Sketch, Studio Portrait, Isometric, Anime, Low Poly, Surreal, and a dozen more. Pick one and both output prompts are built around it. The catch is the first option: Own Style, which routes you to own_style - type your own style description there and the presets stand down.
• max_tokens (default 300, up to 8192) - the response cap you'll request from the LLM. 300 is right for prompt expansion; raise it if your LLM is writing an essay when you wanted a prompt.
• additional_system_prompt - extra instructions merged into the system prompt (formatting rules, "only output the prompt, no preamble," that kind of thing).
• fit_composition_to_style - when on, the detail prompt also forces composition guidance matched to the style. Handy if you noticed your LLM-expanded prompts always forget the framing.

Install

Usual pack story: ComfyUI_StarBetaNodes is the retired beta repo and the nodes ship in ComfyUI_StarNodes now. Install via ComfyUI Manager (search Starnodes) or git clone https://github.com/Starnodes2024/ComfyUI_StarNodes into custom_nodes, then restart. No Ollama install is required by this node - you still need Ollama and a model for the actual expansion, but the node is just prompt text.

Where people get burned

The "JC" in the name isn't a model - it's the author's variant label, so don't go hunting for a JC checkpoint. And the most common mistake is wiring system_prompt into your image encoder instead of your chat node. Read the three outputs, don't just grab the first one: system_prompt → LLM system message, detail_prompt → appended to your generation prompt, style_name → your metadata. The author's own tools run Ollama this way, so the shape of these outputs is battle-tested even if the node itself is a quiet little helper.