Minimax H3 Prompt Engineer · RunningHub

Minimax H3 Prompt Engineer · RunningHub is the legacy prompt-only node from the colorAi pack, and it's the one to reach for when you want the H3 prompt format handled for you but not the whole conditioning stack. It calls RunningHub's LLM API, gets back a properly structured H3 prompt document, runs it through the pack's deterministic validator, and - if you've got auto_repair on - asks the same model to fix any structural mistakes once. What it does not do is touch any model weights. No GPU, no H3 checkpoint, no VAE. It's an output node whose four outputs are all strings. The README is blunt that new workflows should prefer the flagship MiniMax H3 Prompt Studio + Generate node instead, and it's right - but this one still has a real job: feeding a validated H3 prompt into a hosted H3 API, or into the official H3 node as text, without dragging in the full studio.

How it works

This is a straightforward LLM call with a safety net. The node builds a request from your task_mode, user_request, duration_seconds, and reference_context, sends it to RunningHub's OpenAI-compatible chat completions endpoint (llm.runninghub.ai for the Global site, llm.runninghub.cn for China), strips the chat scaffolding from the response, and validates the result. The validator checks the things H3's format is strict about: sequential [Shot N] blocks, MM:SS.mmm cut timestamps that strictly increase, dialogue wrapped in <d>[language] ...</d>, and Full Reference sections. If the first pass is invalid and auto_repair is on, it repairs once with deterministic sampling; if that still fails and strict_validation is on, the workflow stops instead of silently passing garbage downstream.

The inputs that matter

• runninghub_api_key - an Enterprise-Shared key with LLM access, and runninghub_site selects Global or China. The model dropdown is populated live from the chosen site's /v1/models endpoint (62 options on current builds), with a fallback list if discovery fails; custom_model overrides the dropdown if you know the exact ID.
• task_mode - one of T2VA (text to audiovisual), I2VA / FL2VA / L2VA (first/last frame modes), or FULL_REFERENCE.
• user_request and duration_seconds - the actual creative brief and target length.
• reference_context - the one that trips people up. This node can't see your videos or audio, so anything the LLM should know about references has to come through here as text. That's the biggest reason to graduate to the Studio node later.

The optional inputs (creative_presets, first_frame, last_frame, reference_images) feed reference alignment in the prompt - they don't generate anything. temperature at 0.2, max_tokens at 4096, and the two validation toggles on is a sensible starting point.

What you do with it

Wire formatted_prompt (the cleaned, validated H3 document) into wherever your H3 prompt needs to go - the official conditioning node if you're running locally, or a hosted H3 workflow. validation_report tells you if the structure passed, raw_response shows the model's original output (useful when validation fails), and usage_json carries the endpoint, model, request IDs, and token usage for billing sanity.

Installing and gotchas

Same pack, same install: search "MiniMax H3 Prompt Engineer" in ComfyUI Manager, or clone https://github.com/colorAi/ComfyUI-Minimax-H3-Prompt-Engineer into custom_nodes/, pip install -r requirements.txt (just certifi), restart. Because this node makes no LLM call to anything but RunningHub, it's the one node in the pack that works in regions where the H3 weights are license-locked (the US, EU, UK, South Korea) - a prompt is just a prompt. Two habits worth keeping: model availability varies by site, so if the dropdown looks empty check the site's /v1/models endpoint is reachable, and clear runninghub_api_key from your workflow before sharing it, since ComfyUI can serialize widget values into the JSON. The key is the kind of thing the KB keeps warning about: a node whose whole job is phoning home with a credential, so treat it accordingly.