MiniMax H3 Prompt Preview & Edit

There's a limbo moment in every LLM-driven video workflow: the director model just spat out a five-paragraph MiniMax H3 screenplay, and you'd really like to read it before it costs you a render. That gap is exactly what MiniMax H3 Prompt Preview & Edit fills. It sits between the prompt writer and the sampler as a syntax-aware inspection point - the node you glance at, fix, and forward.

It's the younger sibling of the same pack's H3_Promptor (the LLM screenwriter) and pairs naturally with H3_PromptComposer for hand edits. MiniMax H3 is the open-weights 33B omni video model ComfyUI supported from launch (minimax-h3.json has the model rundown); it's famously picky about its prompt structure - shot headers, subject tags, section blocks - so a human review pass before sampling is less a luxury than a survival habit.

How it works

Mechanically it's a pass-through with a memory. execute() takes whatever text arrived on its input socket and returns it unchanged as a string - but if nothing arrived, it falls back to a stored copy of your last edited text. That's the _stored_prompt widget: internal state (the underscore tells you not to wire into it) that keeps your edits alive across a workflow save and reload. Like its sibling, it returns float("nan") from IS_CHANGED so it always fires and pushes fresh text downstream on every queue.

The value is again in the front end, which renders the H3 prompt with the same cinematic syntax highlighting as the composer - shot blocks, <Subject N> tags, section headers and dialogue all colour-coded. From there you can:

• Edit inline, or highlight one shot/subject/snippet and hit ✨ Refine to have a configured LLM polish just that piece without rerunning the whole expensive writer workflow. Structural tags and timestamps are locked while the LLM works, so it can't corrupt the schema it's improving.
• A/B your way out of a mistake. Refined results preview before applying, and after apply a gold ↩ Restore Original button reverts the whole prompt in one click. Editing a 1,500-word cinematic prompt has never felt this risk-free.

Inputs and output

Keep it simple: the prompt input (optional) is where you connect the PROMPT output from MiniMax H3 Promptor - the tooltip says exactly that. Leave it unconnected and the node just shows whatever you last had stored, which is also handy as a plain, highlighted preview of any H3-syntax string. Output is a single PROMPT STRING, your original or edited text, ready to feed the same downstream H3 sampler the writer node fed.

Install

Same pack, same story: ComfyUI Manager → search "MiniMax H3-Promptor", or:

cd ComfyUI/custom_nodes
git clone https://github.com/1038lab/ComfyUI-MiniMax-H3-Promptor
pip install -r requirements.txt

Restart and you're done - dependencies are just requests, no weights to haul. The Refine modal needs an LLM provider configured under the gear Settings → MiniMax H3 (Ollama and LM Studio work locally and are on by default).

Where people get burned

Two habits trip users up. First: if you disconnect the input wire after editing, the node quietly reverts to its stored copy - the socket wins whenever it's connected, so leaving a stale wire attached can resurrect old text when you expected your edit. Second, keep the shot count honest: if you delete a [Shot 2] block here, make sure the video length set on the sampler still makes sense, because nothing downstream checks. Preview the prompt, fix the shot that reads wrong, and let the sampler do its thing - that's the whole job, and it's a genuinely useful one.