T8 Prompt Inspector (Local, Non-blocking)
A cheap sanity check before you spend credits on a bad prompt
- original_prompt
- warnings_json
- summary
Every video-generation API bills you whether the prompt was good or broken. T8 Prompt Inspector is the five-cent insurance policy that runs before you spend: it reads a prompt that's supposed to be in MiniMax H3, Seedance 2.0, or Music 3 format and checks whether the structure is sound. Completely local, completely deterministic, and it never calls an LLM - which is exactly why it's safe to drop into any workflow and leave there.
It's "non-blocking" on purpose. The inspector does not veto anything, it doesn't short-circuit your graph, and it never rewrites your prompt. It hands your text straight through and reports what it found. The upstream node runs regardless; the warnings are for you, not for ComfyUI.
How it works
Feed it the output of one of the three T8 enhancers (or any hand-written prompt in the same format) via the prompt input. prompt_family defaults to AUTO, which sniffs the format: it's Music 3 if it sees ### Global Metadata / ### Vocal Details / ### Arrangement headers, H3 if it sees the H3 fields or [Shot N] markers, otherwise Seedance 2.0. You can force a family if the detector guesses wrong.
The checks are the boring, mechanical kind that LLMs are bad at and regex is good at:
[Shot N]/镜头Nnumbers must start at 1 and climb consecutively - a jump from 2 to 5 is the classic hand-edit tell.expected_shot_countandexpected_languagelet you assert "this should have 8 shots / be in English" and get a warning when reality disagrees.instrumentalflags a Music 3 caption that mentions a singer when it shouldn't.duration_secondscross-checks against what you told the generator.
Everything else about creative quality is out of scope by design - the tooltip on the node says it plainly: it checks re-computable structure, not whether the prompt is any good. That's the honest boundary.
The three outputs
original_prompt- your text, unchanged, verbatim. This is the node's pass-through so it can sit inline in a chain.warnings_json- the machine-readable list, each with a code, severity and message.summary- a human-readable STRING of what it found, for a T8 Show Text or any text preview node.
If you just want the checks, ignore original_prompt and read summary; the JSON output exists for the kind of person who parses things.
Install
Same pack as the enhancers - 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.git
Restart, Ctrl+F5. There's a bundled example workflow (prompt_inspector_local_qwen_example.json) showing the sensible setup: H3 enhancer → inspector → check the summary before the prompt goes anywhere expensive.
Where people get burned
- Expecting it to judge quality or fix problems. It won't; it flags structure and moves on. That's the feature.
- Wiring it somewhere that requires the prompt to be transformed. It passes
original_promptthrough untouched, so use that output if you're chaining, not the summary. - The AUTO family detector is a heuristic - if you're feeding it something genuinely hybrid, set
prompt_familyexplicitly rather than trusting the sniff.
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| prompt | STRING | — | |
| prompt_family | COMBO | AUTO(本地识别) | 4 options: AUTO(本地识别), MiniMax H3, Seedance 2.0, MiniMax Music 3 |
| expected_shot_count | COMBO | AUTO | 21 options: AUTO, 1, 2, 3, 4, 5, +15 |
| expected_language | COMBO | AUTO | 3 options: AUTO, 中文, English |
| instrumental | COMBO | AUTO | 3 options: AUTO, 是, 否 |
| duration_seconds | INT | 00–900 | — |
| task_intentopt | STRING | — |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| original_prompt | STRING | — |
| warnings_json | STRING | — |
| summary | STRING | — |