Task System Prompt Selector 👻

Every LLM-based prompt node in this pack needs a system prompt that says "you are a specialized text-to-video assistant," and every one of those is a sentence you've typed before. Task System Prompt Selector 👻 is the pack's answer to that redundancy: a dropdown of 13 task roles, each matching ByteDance's Bernini prompt-engineering conventions, and it hands you the right system prompt plus a ready-to-send system,user combined string.

If you've never heard of Bernini - it's ByteDance's video generation/editing model family, fine-tuned from Wan weights and released Apache 2.0. Its prompt_enhancer defines a clean taxonomy of generation jobs (text-to-image, text-to-video, image-to-video, video-to-video, reference-based, ad insertion...), and this node is a faithful port of that taxonomy. You don't need Bernini to use it, though - a system prompt is a system prompt, and these roles work with any LLM-in-the-graph workflow.

The two inputs

• task_select - the dropdown. 13 options like [t2i] 文本描述生成单张图像 (text-to-image), [t2v] 纯文本指令从零生成短视频 (text-to-video), [i2v], [v2v], [r2v], [ads2v] (ad insertion), and a [default] fallback that just says "you are a helpful assistant". Each label is [task code] + plain-language description, which is a nice touch - you pick by what you're doing, not by remembering codes.
• user_prompt - your actual instruction, multiline. This gets appended to the chosen system prompt.

The two outputs are the whole point:

• system prompt (STRING) - just the role text, so you can wire it into anything that wants a standalone system prompt.
• prompts (STRING) - the concatenated system,user pair (English comma between them), ready to drop into a chat-message node.

Why you'd bother

Two reasons, and they're both real. First, the system prompts are actually good - each one is scoped to its task ("You are a helpful assistant specialized in text-to-video generation" vs "...specialized in image editing"), and the LLM-in-the-graph crowd has learned that a tight, task-scoped system prompt drifts less than an open-ended "rewrite my prompt" instruction. Second, it kills the copy-paste: you swap a dropdown, not a text block, and the combined output means one wire into your LLM node instead of assembling the pair yourself.

The catch: it's not magic. The system prompts are single-line role definitions, not elaborate few-shot scaffolding. If you want a real prompt-enhancer (something that rewrites, expands, adds shot structure), that's the pack's other nodes - MiniMax H3 Prompt Optimizer, which builds on this exact idea with far more machinery. This node is the light version: correct role, every time, zero effort.

Install

# ComfyUI Manager: search "ComfyUI-Minitools"
# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/lepiai/ComfyUI-Minitools
# restart ComfyUI

No dependencies beyond the standard library - it's a dict lookup with string concatenation. It doesn't call any API and needs no key; it's purely a text presets node.

Gotchas

• The labels are bilingual-Chinese-first. Task descriptions are in Chinese with English task codes. If you can't read Chinese, the codes (t2i, v2v, rv2v...) are clear enough, and the English system prompt that comes out is what actually gets used.
• The combined output is naive. It's literally f"{sys_prompt},{user_prompt}" - English comma. If your downstream LLM node expects JSON or a newline separator, you'll want to take the two outputs separately and format them yourself. The system prompt output exists precisely for that.
• It's tied to Bernini's task taxonomy, which is video-generation-centric. Fine if you're in that world; a generic image-only workflow will mostly live on [default] and [t2i].

Small, boring, dependable - which is exactly what you want from a presets node. It won't change your life, but it removes one repetitive step and makes your LLM nodes consistent. Sometimes that's the whole win.