H3 Final Prompt Editor

H3 Final Prompt Editor is the exit door of the H3 Scribe pack. Everything upstream - the Qwen analyses, your Japanese authoring, the Compose pass - funnels into a single English H3 prompt, and this is the big editable text box where you see it and can still change it before generation. It's the "phase-2 partial-execution target" of the workflow: the final, human-reviewable state between Compose and the MiniMax H3 generation nodes.

You might wonder why you'd want to hand-edit a machine-built prompt at all. Because H3 prompts are long and structured, and sometimes you just want to fix a phrase, tighten a shot description, or paste in a camera move the composer got wrong - without tearing up the whole authoring state and re-running both Qwen passes. This editor is the cheap way to do exactly that, and it's designed so the cheap way doesn't cost you the cache.

The persistence trick

Like the Authoring Editor, this node is built on a snapshot merge, but it's even simpler - deliberately so. It takes three inputs: text (your current text), source_snapshot (the Compose output it last saw), and source (the fresh Compose result). The logic is a single rule: if the upstream Compose output changed, the editor refreshes to the new text; otherwise your current text is returned verbatim, untouched.

That "verbatim" is load-bearing. The node's comment says it plainly: it deliberately does not strip, normalize, render, or re-compose the text. Whatever you typed is what reaches the generator, byte for byte. No surprise whitespace trimming, no sneaky re-render from a stale state. And because the editor is an output node, ComfyUI's caching means editing only this node invalidates the downstream generation path - the upstream Qwen results stay cached, so you don't re-analyze and re-compose images just to tweak one sentence. The web-side JS persists your text in the node's properties so it also survives saving and reloading the workflow.

Interface

One required text input, one required source_snapshot, one optional source - in the bundled workflows they're all wired so that source comes from the Validate & Render node. The single output, also named text, feeds your MiniMax H3 generation graph. If you run it before anything upstream has produced a prompt, you'll just see an empty box; there's no error, because unlike the Authoring Editor there's nothing to validate here - it's a dumb, faithful text holder, and that's its job.

Install

The whole pack installs via ComfyUI Manager (search "H3 Scribe") or git clone https://github.com/last-git/h3_scribe into ComfyUI/custom_nodes, then restart. Requirements beyond that: a recent ComfyUI with the native MiniMax H3 nodes, ComfyUI_Simple_Qwen3-VL-gguf for the Qwen side, and the model files from the README. The pack itself only depends on pydantic and makes no API calls.

One habit worth forming: after a re-analyze changes your subject lineup, re-check this box before generating. Compose can hand you a perfectly valid prompt about the wrong person if you edited the authoring but never re-ran Compose - the editor happily shows stale-but-consistent text until its upstream changes. The node won't warn you; it trusts its snapshot logic. That's the trade you accept for the cheap-edit path.