T8 Directed Revision（定向修改）

You know the feeling: you're three hours into a video workflow, the prompt is basically perfect except one scene, and you paste it into a chat LLM asking for "a small tweak" - and it hands you back a completely different prompt that lost your character's jacket, the reference label, and the exact dialogue. T8 Directed Revision（定向修改） is the node built specifically to stop that from happening.

It makes exactly one LLM request, but the request is written like a surgical-editing contract: modify only the scope the user named; preserve every locked anchor, unmentioned identity, costume, prop, dialogue, lyric, visible text, reference label, timing fact, and model-native field; never restart from scratch. The whole prompt-engineering move of this pack's author, T8 (T8mars), is giving the model a tight job description instead of hoping. If you've read the KB's material on LLM prompt-enhancer drift - where an unconstrained enhancer quietly invents narrative you never asked for - this node is the fix applied to the revision step specifically.

The inputs that matter

• original_prompt (required) - the prompt you're revising. No original, no revision.
• revision_request (required) - what to change. Be specific: "swap the second shot from a dolly-in to a static close-up," not "make it better."
• locked_anchors - one per line, things that must survive verbatim. This is the insurance policy; the system prompt says to preserve them and the node locally warns if they vanish.
• creative_brief - optional, from the T8 Creative Director. When connected it outranks stylistic enrichment, so a locked identity in your brief can't be quietly re-imagined.
• rewrite_mode (strict / balanced / creative) and output_language (中文 / English), plus the usual seed.

On the paid-request side, the node shares the pack's provider plumbing: default is the Seedance NZ gateway, or wire a T8 LLM Provider Config node's output into provider_config to use the AI Workshop, an OpenAI-compatible endpoint, or a local Qwen GGUF with no key. api_key accepts a plain STRING if you're not using the shared config.

What comes out

• revised_prompt - the new text, ready for your H3 or Seedance enhancer.
• revision_report_json - structured summary: what changed, what was preserved, warnings. Marked with structured_response so you know whether the model returned clean JSON.
• unified_diff - a local diff between original and revised, generated on your machine. This is quietly the best feature: you can see exactly what moved, line by line, without trusting the model's own summary.

Installing

In the comfyui-minimax-h3-prompt-enhancer-T8 pack - Manager search "MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8)", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-prompt-enhancer-T8

Restart and Ctrl+F5 afterward. Cloud mode adds no dependencies; it rides on ComfyUI's bundled requests, NumPy and Pillow.

Gotchas worth knowing

Both required fields must be non-empty or it errors up front - good behavior, it fails before spending tokens. The "one request" promise is real: there's no second pass to fix formatting, and if the model returns non-empty but non-JSON text, the node keeps the raw text rather than discarding a valid revision. And remember the pack's honesty boundaries: this node only rewrites text; it doesn't watch a video or check a render. One small thing that's easy to miss - because the local diff is generated from the returned text, it's a diff of prompts, not of rendered footage. It tells you what changed in the words, which is exactly the thing you asked to control.