MiniMax H3 Prompt Guide Builder

Every H3 enhancer node in this pack has a secret: the "brains" are just a very long system prompt describing MiniMax H3's documented prompt contract, plus a user prompt built from your settings. This node hands you that system prompt and user prompt as plain text, without running any LLM at all. If you already have a text-generation node you like - Ollama, LM Studio, a GGUF loader, or the QwenVL Prompt Enhancer the description name-checks - you can wire those outputs into it and effectively build your own enhancer around your model of choice.

Why this exists

The enhancer nodes in the pack are opinionated: they pick the backend (API endpoint or managed llama-server) and drive the whole thing. That's convenient, but if you want your own model, your own server, or your own front end, you don't need the scaffolding - you need the instructions. The Guide Builder is the model-free core of the pack, exposed as a node. Its job is deterministic: take your basic_prompt, mode, duration_seconds, reference_context, and the pile of style/audio/resolution widgets, and compile them into two strings.

• system_prompt - the full H3 rewriting contract: the strict 6-block anatomy for Ref2VA, the 3-block structure for T2VA/I2VA/FL2VA/L2VA, verbatim <d>[Language] ...</d> dialogue rules, the multilingual/dialect table, physical-action chains, the "name what a camera could record" rule.
• user_prompt - your request formatted for the model: the source text, resolved mode, duration, reference context, and every policy you toggled.

Pipe those into any chat-completions text node, feed the reply through a MiniMaxH3PromptValidator to check the result, and you've rebuilt the enhancer with your own engine.

Inputs worth knowing

mode (default auto) resolves the generation type; the node reports what it decided through the resolved_mode output so you can see whether auto guessed right before you send anything downstream. creative_latitude is the same four-level dial as the enhancer nodes - how far the writer may go beyond your text, from verbatim_source (reformat only) to invented_production (build the world). The audio trio is worth a glance: ambience_foley_policy (auto/ensure_audible/off), background_score_policy, and instrumental_description - describe concrete instrumentation and tempo, because mood words get translated into audible parameters anyway. dialogue_language (default auto) sets the language your quoted dialogue stays in.

Two underrated details. First, lora_trigger_words are appended to the finished prompt verbatim, after validation - a trigger token like g0r3_style is an exact token the LoRA was trained on, so the node keeps it out of the LLM's translator hands. Second, the resolution math happens here too: aspect_ratio plus target_megapixels produce the width/height outputs (16-pixel aligned), so you can drive your latent size even in a custom pipeline.

Outputs

system_prompt, user_prompt, resolved_mode, treatment_warnings (drops and conflicts between your treatment settings), and width/height. If you're only here to copy the spec into your own prompt-writing setup, system_prompt is the whole ballgame.

Install

Same one-liner as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/hyukudan/ComfyUI-MiniMax-H3-Prompt-Enhancer.git

Then restart ComfyUI (or search ComfyUI-MiniMax-H3-Prompt-Enhancer in Manager). This node needs nothing beyond ComfyUI's standard environment - no LLM, no API key, no model files. It's pure string compilation, so it works fully offline on day one.

The gotcha

The Guide Builder is only as good as what you feed it into. It emits instructions, not a result - if your downstream LLM is weak at following long structured formats, or isn't an instruct-tuned chat model, the output will be chatty prose with preamble, and you'll see the validator fail on structure. That's not this node's bug; it's the classic LLM-as-tool failure mode from the KB: a chat model doesn't emit clean structured output by default. Chain the validator after it, and if the model keeps talking back, that's your signal to switch engines.