MiniMax H3 Rich Scene Prompt Editor (Experimental)

Rich Scene Prompt Editor is the fancier sibling of the Scene Prompt Editor, and it's labeled experimental for good reason. Where the plain editor gives you room to type, this one goes further: color-coded references so you can see @tag and #semantic[timestamp] anchors at a glance, media previews inline, prompt guides, revision history, and an optional one-click agent rewrite powered by a Direct API or MCP configuration you set in ComfyUI Settings.

Let's be clear about the "experimental" part before you build a production around it: this node changes only the selected scene's prompt in the upstream Plan. It does not edit Plan settings, schedules, or seeds. It's a prompt-writing surface with training wheels on top of the reference system, not a Plan editor.

What it does

Connect the Plan's plan output. The editor shows the active scene's prompt with its references rendered as chips - click a chip to swap that exact occurrence for another compatible connected reference, toggle a tagged picture between native @tag and semantic #tag[time] syntax, or edit the semantic timestamp in its own field without hand-editing the prompt. That's the genuinely useful part: reference management is usually buried in a registry, and here it's inline where you're already writing.

The revision history is quietly valuable for the same reason version control is: when you queue a scene, dislike the result, and want to know what changed from the take that worked, the history answers it.

The agent rewrite is the feature that will either sell you on this node or make you turn it off. Configured via ComfyUI Settings (a Direct API key or an MCP agent), it can rewrite the active prompt on demand. Some people love it for expanding a terse idea into a full H3 prompt; others find a model rewriting its own conditioning prompt a bit too much, especially since H3 is an omni-modal model that reads the prompt and the references. Either stance is defensible. The node works fine with it disabled.

At execution time the plan output passes the connected Plan through unchanged - like the plain editor, this is authoring-side sugar. You can put it on an editor-only branch and never affect the generation graph.

Install

Part of ComfyUI-MiniMaxH3-Contex-Loop - ComfyUI Manager → search the pack, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart. No pip dependencies; the agent rewrite needs whatever API/MCP backend you configure in ComfyUI Settings. Generation-side, it expects a current ComfyUI with native Add Guide for MiniMax H3, and the H3 weights aren't bundled (geofenced out of the US/EU/UK/South Korea under the Community License, in case you're in those regions).

Gotchas

Don't reach for this node expecting it to fix a seed or a duration - it can't, and it says so. And remember the golden rule that applies to every prompt-editing node in this pack: an edit to a scene that's already been checkpointed is a change that invalidates dependent resume. If you're mid-production and a later scene depends on scene two, editing scene two's prompt here and re-queuing will trip the resume validation. That's the pack protecting you from mixing runs, and it's exactly the behavior you want once you've been burned once.