LocalLLM System Prompt (Library)

This node does one thing: it hands out a string. No model, no VRAM, no generation - pick a template, get a system_prompt, wire it into the system_prompt_input socket of the Prompt Generator or the Image Caption node. Weirdly, it's the node most people skip for months and then wonder how they lived without.

Here's why it earns its slot. A good system prompt is not a sentence - it's a page of rules: how to structure the output, what to never mention, how to handle text rendering, which token to wrap the answer in. Krea 2 users trade these files on the subreddit like presets. Pasted into a widget, that page lives in one workflow and dies with it. Put it in a node, and every graph points at the same file.

How it works (it's a folder)

At node-definition time the node scans prompts/*.md inside the pack. Each file's first # heading becomes the dropdown name, its filename slug gets appended in brackets, and the first fenced code block is the prompt body - files with no code fence are used as-is. README.md is skipped. That's the whole mechanism, so the dropdown is exactly a listing of that folder.

Nine templates ship. Three are for image models (Krea2 / FLUX2 / FLUX2 Klein Base): plain image captioning, captioning with a person swap, and an idea-expander that lets you describe people freely. Six are for MiniMax H3 video: text-to-video, image-to-video, and four reference-to-video combinations (person images, background image, one or two voice tracks).

Inputs and output

• template - the enum above.
• append - multiline, optional extra instructions glued onto the end of the template with a blank line. This is the field you actually use: keep the library template intact and append your house style, your character sheet, or "no text in the image".
• Output system_prompt (STRING) → system_prompt_input on either main node.

And note which one wins: on both consumer nodes, a connected socket overrides the system_prompt widget entirely. The tooltip says so, but people still edit the widget for ten minutes wondering why nothing changes.

Adding your own

Drop an .md file into prompts/, give it a # Title and a fenced block, then press R in ComfyUI (refresh node definitions) and restart-free it appears:

# My House Style

Always output the answer inside <prompt>...</prompt>.
Never mention camera gear. Comma-separated, one line, under 60 words.

(The # My House Style line must be a real heading, and the instructions above should sit inside the file's first fenced code block - that's what the parser reads.)

That <prompt> convention is not decoration - the Prompt Generator returns the contents of the last <prompt> block, which is how the answer arrives without preamble attached.

Two things to know before you use the video templates

They speak German on purpose. The H3 video prompts hard-code [German] for dialogue, which is a convention of this library, not an H3 requirement. Replace every [German] with [English] - the English language name, capitalized exactly like that. [Deutsch] and [GERMAN] are not it. H3 supports eleven dialogue languages.

They assume headroom. The library's own readme says its prompts want n_ctx 8192, max_tokens 4096, suppress_thinking on, and temperature 0.5–0.6. Feed them a smaller context and your video prompt comes back cut off mid-sentence, which reads like a bad model and is really a bad budget.

One failure mode, from the source: if the template you saved in a workflow no longer exists - you renamed or deleted the file - the node raises Template 'X' not found in ... Press 'R' to refresh. Re-pick it in the dropdown and it's fine.