MiniMax H3 Prompt Lint / 提示词检查

MiniMax H3's prompt format is unforgiving, and worse, it's quietly unforgiving. Misspell a field name, mis-time a shot cut, forget a speaker ID, and H3 won't error out - it just produces a worse video and never tells you why. The Prompt Lint node is the deterministic check that catches those mistakes before you spend fifteen minutes rendering.

What it is (and isn't)

It's a linter, not a generator. There's no LLM inside, no rewriting, no API key. It takes your prompt, validates it against the rules in MiniMax's own video prompt writing guides (the base guide for t2va/fl2va modes, the reference guide for ref2va), and passes the prompt through unchanged. That's a deliberate design choice - the pack explicitly doesn't want to duplicate prompt-generation/rewriting nodes like boo-textgen or PromptEnhancer. It only verifies.

Because the prompt passes through untouched, the node can sit inline in a workflow: lint result on one branch, clean prompt fed forward on the other.

What it checks

This is where the node earns its keep. Set mode to t2va, fl2va, or ref2va and it validates the fields that mode requires (three base fields for t2va/fl2va, six full-reference fields for ref2va), then digs into the details:

• required fields present and in order, none empty
• shot numbering sequential from 1, cut timestamps in the MM:SS.mmm format, strictly increasing, and Shot 1 carrying no timestamp at all
• dialogue tags with a recognized language tag from the 11 that have stable support, balanced <d> tags
• speaker IDs consecutive from S1, and dialogue attributed to a speaker in its shot
• voiceover using the required exact phrase "says in an off-screen voiceover" plus a lips-closed statement
• soundscape within 1–4 sentences and not restating dialogue; music within 1–3 sentences and free of diegetic words (radio, TV, jukebox…)
• in ref2va mode: task-type prefixes, the fixed preservation markers, and <Picture/Video/Audio/Subject N> numbering within what's actually connected

The inputs that matter

prompt (multiline), mode, and duration_seconds (set the clip length and it flags cut times past the end; 0 disables that check). If you're in ref2va mode, also set ref_images, ref_videos, ref_audios to how many references are actually connected, so out-of-range <Picture N> tags get flagged.

Outputs: prompt (unchanged), warnings (a numbered list, or "clean - no issues found"), and is_clean (a boolean). Wire warnings into a text preview and is_clean into whatever you want to gate.

The one setting to know: raise_on_warning. Default off, so a lint warning never blocks a render - it just reports. Flip it on and a warning fails the workflow, which is handy once you've got a template you want enforced on every run.

Install

Part of the MiniMax-H3 Tools pack (ComfyUI 0.30.0+). Search "MiniMax H3 Tools" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Rinne414/ComfyUI-MiniMaxH3-Tools.git

Restart ComfyUI, find MiniMax H3 Prompt Lint under MiniMax H3. No extra dependencies, no model files - it's pure string validation.