minimax-h3

Adapt Story-Film shot briefs into MiniMax H3 audio-video prompts using the mandatory H3 prompt-writing format plus an appropriate optional style skill for T2VA, I2VA, FL2VA, L2VA, or Ref2VA.

SKILL.md

GitHub

MiniMax H3 Prompting

Required skill stack

Every MiniMax H3 prompt uses this stack in order:

  1. Story-Film canon, continuity, reference authority, approved dialogue/audio, and the source shot brief.
  2. The selected ComfyUI workflow, which determines the H3 mode and actual runtime capabilities.
  3. skills/h3-prompt-writing/SKILL.md, which is mandatory and controls final H3 field names, keyframe alignment, shot timing, speaker/dialogue syntax, and base-versus-Ref2VA section order.
  4. Zero or one automatically selected H3 style skill from references/MINIMAX_H3_SKILL_ROUTING.md. Use more than one only when the user explicitly asks for a hybrid.
  5. This Story-Film adapter's project-specific continuity, exact-audio, reference-scope, and capture constraints.

Story-Film runs python scripts/minimax_h3_skill_router.py --text "<brief>" automatically whenever the selected prompt adapter is minimax-h3; the user does not need to call h3-prompt-writing or a style skill separately. An explicit user selection of one of the eight style skills overrides automatic style classification but still returns through h3-prompt-writing, this minimax-h3 adapter, deterministic H3 format validation, and prompt-qc.

A style skill enriches content. It never replaces h3-prompt-writing, never changes the selected workflow, and never overrides Story-Film canon, reference authority, approved-audio authority, or workflow/runtime capability.

Prompt audit metadata

Store these values in the Story-Film prompt file outside the final model prompt:

adapter: minimax-h3
h3_base_skill: h3-prompt-writing
h3_style_skill: <selected-style-skill-or-none>

The generation operator sends only the final H3 prompt body to the model.

Select mode

  • T2VA: text to audio-video
  • I2VA: start from a supplied first frame
  • FL2VA: connect supplied first and last frames
  • L2VA: converge to a supplied last frame
  • Ref2VA: full-reference generation using labeled image, video, and audio references

Use the mode actually implemented by the selected workflow. Do not infer FL2VA/L2VA support merely because a last-frame asset exists.

Base mode output

Read skills/h3-prompt-writing/references/base-en.txt.

Preserve this core order:

  1. integrated_multimodal_description
  2. overall_soundscape
  3. non_diegetic_music

I2VA, FL2VA, and L2VA also require the correct alignment instruction before the core fields. Build a chronological audiovisual timeline. [Shot 1] has no cut timestamp; later cuts use strictly increasing At MM:SS.mmm times.

Ref2VA output

Read skills/h3-prompt-writing/references/ref-en.txt.

Preserve this order:

  1. subject_definitions
  2. summary
  3. retention_analysis
  4. detailed_description
  5. overall_soundscape
  6. non_diegetic_music

Use stable <Subject N>, <Picture N>, <Video N>, and <Audio N> labels. State exactly what is retained from each reference and where it appears.

Story-Film rules

  • Preserve dialogue, lyrics, and visible scene text exactly in the requested/source language.
  • Use stable (S1), (S2), and later speaker IDs and H3 <d>[Language] ...</d> blocks.
  • When the source brief requires visible-dialogue sync, preserve exact speaker, mouth visibility, cut policy, and measured speech duration.
  • Keep reference labels consistent and preserve Story-Film reference authority scopes. A character atlas is not automatically composition authority and a temporal tail is not identity authority.
  • Tie first or last frames to the timeline explicitly in keyframe modes. When end_frame.required is present, end the described timeline in that state and use FL2VA/L2VA only when the selected local workflow actually uses last-frame conditioning.
  • For Ref2VA continuation, a registered visual-only previous-shot tail may carry temporal continuity. Its audio must be stripped before it becomes a visual reference.
  • When approved dialogue audio exists, preserve model-neutral seconds, approved media identity, and the exact approved source. If the selected workflow uses MiniMaxH3ExactAudioLock, that workflow is generation authority for exact target audio: convert approved start seconds to H3's 24 fps target timeline, feed each independently timed source through MiniMaxH3TimedAudio, lock the resulting H3 target-audio latent with MiniMaxH3ExactAudioLock, and when the selected workflow includes it, reinforce the same source through core MiniMaxH3AddGuide at the same target frame.
  • A prompt/style skill must not replace that path with H3-regenerated dialogue/music, prompt-only timing, post-hoc muxing as a substitute for conditioning, a different workflow, or a reduced speaker/event model. MiniMaxH3ExactAudioLock uses autogrowing timed-audio inputs; Do not impose an artificial speaker or utterance count. Preserve the workflow's selected overlap, overflow, gain, and exact-audio policies.
  • If a required exact-audio node is absent from live /object_info, block the selected workflow and report the missing capability. Do not silently install the node and do not silently choose another workflow.
  • Preserve source capture_behavior as visible capture properties rather than inventing camera hardware.
  • Use concrete image and sound descriptions instead of generic quality adjectives.
  • A style overlay must not invent brand facts, dialogue, lyrics, scientific claims, character identity, reference roles, or continuity facts.
  • Save under 04_generation/prompts/minimax-h3/<shot-id>.md.

Deterministic format gate

Before prompt-qc or generation, run python scripts/minimax_h3_prompt_validator.py --mode <T2VA|I2VA|FL2VA|L2VA|Ref2VA> --duration <seconds> --prompt-file <final-model-prompt-body>. Pass only the final H3 model-prompt body, not Story-Film audit metadata. A validation error is a blocker. Repair the prompt and rerun the validator; do not submit malformed H3 syntax to ComfyUI.

Done

The prompt file records its H3 base/style routing, all references resolve, the deterministic H3 format gate passes, H3 section order and syntax are correct, and described audiovisual time equals the requested duration.