h3-guide

H3 Guide

Turn vague ideas into an executable MiniMax H3 plan, and run it through the existing opc skill. This skill handles creative decisions, asset roles, test matrices, and QA; the final structure of the Prompt is handled by the official MiniMax h3-prompt-writing skill.

Prerequisite rules

1. When creating, rewriting, or expanding a Prompt, first read the official h3-prompt-writing skill in full; in Base mode, read its references/base-en.txt, and for Ref2VA read its references/ref-en.txt。
2. When actual video generation or editing is needed, then read the available opc skill and its references/video.md, subject to the current CLI capabilities.
3. First inspect the images, videos, and audio provided by the user. For videos, at least read duration, dimensions, frame rate, and audio track, and extract frames for observation; assign roles only after actually viewing the images.
4. When the user only requests planning or a Prompt, do not submit a generation task. When the user requests generation, testing, or continued iteration, complete safe generation, QA, and delivery.
5. Do not treat style words as a plan. Clarify the story or function, time beats, reference asset responsibilities, camera, sound, invariants, and prohibited items.

Prompt format ownership

OPC workflow	Official mode	Final structure
h3-t2v	T2VA	integrated_multimodal_description、overall_soundscape、non_diegetic_music
h3-i2v + first frame	I2VA	Official first-frame alignment sentence + Base three fields
h3-i2v + first and last frames	FL2VA	Official first-and-last-frame alignment sentence + Base three fields
h3-i2v + last frame	L2VA	Official last-frame alignment sentence + Base three fields
h3-r2v	Ref2VA	subject_definitions、summary、retention_analysis、detailed_description、overall_soundscape、non_diegetic_music

The final Prompt uses English; dialogue, lyrics, and on-screen text retain their original language. Official fields, order,[Shot N], timestamps,<Subject N>、<Picture N>、<Video N>、<Audio N>、(S1) and <d> syntax must not be separately invented or rewritten by this guide.

Workflow selection

Goal	Workflow	Selection conditions
Generate a complete scene from scratch	h3-t2v	No fixed character, product, or composition required
Fixed first frame, or specified first and last states	h3-i2v	Composition/subject appearance is determined by the first frame, or the last frame must be reached
Multi-image joint reference, motion/camera reference, video editing, sound reference	h3-r2v	Any image, video, or audio needs to serve as an independent condition

By default, use two-stage generation. For gacha, Prompt testing, and seed screening, first do a low-resolution 5-second Turbo 8 steps preview; after the plan is determined, keep the Prompt, reference asset order, and seed unchanged, remove --turbo, and use Base 20 steps for final output at the user's target resolution. Turbo is a preview tool and is not used as final delivery. At 24 fps, H3's frame grid will output a 5-second request as about 5.167 seconds; when the user specifies other specs, follow the user and use opc video info to verify current support.

Standard process

1. Break requirements into constraints

Extract and confirm:

• Publishing goal, audience, aspect ratio, duration, resolution.
• Subject, scene, action, narrative or functional demonstration.
• Identity, product, composition, camera position, sound, and text that must be maintained.
• The parts that can change, and the parts where change is explicitly prohibited.
• Success criteria: what the audience should feel, remember, and believe after watching.

2. Assign a single responsibility to each reference asset

Write the asset mapping table first, then write the Prompt. Each asset should have as much as possible only one primary responsibility:

• Identity image: face, hairstyle, body type, clothing, or character design.
• Product image: geometry, material, Logo, structure, and color.
• Scene image: space, color, lighting, and art direction.
• Style image: typography, material, visual language; do not let it override identity.
• Storyboard image: shot order and composition changes.
• Video: action, performance rhythm, camera movement, editing, or original scene.
• Audio: timbre, dialogue, music structure, or ambient sound.

h3-r2v The media index is determined by the order in which it appears on the command line, and the Prompt must use exactly the same <Picture 1>、<Video 1>、<Audio 1>. For the upper limit and audio track numbering rules, refer to opc current documentation.

3. Write the Prompt according to the official skeleton

First organize the output contract, global invariants, reference bindings, per-second actions, camera, visuals, sound, text, and prohibitions into creative constraints, then follow the official h3-prompt-writing corresponding mode into the fixed fields. For the complete creative breakdown method, see prompt-framework.md, but none of its local templates may override the official field names and order.

4. Choose the style grammar

Do not just write "cinematic" or "premium." From style-cookbook.md select the corresponding composition, camera, material, editing, sound, and counterexamples. When reusable visual references are needed, first read reference-library.md, then use view_image to view specific assets.

When it involves aesthetic films, TVCs, macro, material transitions, live-action integration, or the user requests "stunning," you must also read aesthetic-mechanics.md. Before submitting the generation, rewrite the aesthetic goals into observable optical, material, spatial, transition, and hold mechanisms; "cinematic," "premium," and "silky" cannot be execution instructions on their own.

5. Execute the editing strategy

For character/object replacement, background and lighting, action/camera movement transfer, multiple simultaneous modifications, and dialogue/timbre editing, handle according to editing-playbook.md . The editing Prompt must first declare the original video's invariants, then list the only parts allowed to change.

6. Testing and iteration

First use --turbo to generate candidates and filter Prompt/seed, do a six-panel frame extraction and key local comparison, then decide whether to retry. After selecting a plan, you must use the same Prompt, reference asset order, and seed, remove --turbo and rerun Base 20 steps, and complete the final quality check with the Base final cut. According to qa-iteration.md determine whether to change the Prompt, swap references, split the task, or acknowledge the model boundary; do not declare success based only on the first frame.

# 抽卡：快速验证创意、Prompt 和 seed
opc video -w h3-t2v --prompt-file prompt.txt \
  --width 608 --height 352 --duration 5 --seed 42 --turbo

# 最终出片：复用选中的 Prompt 和 seed，移除 --turbo
opc video -w h3-t2v --prompt-file prompt.txt \
  --width 1376 --height 768 --duration 5 --seed 42 --steps 20

h3-r2v --turbo It can be used for editing plan previews, but you must focus on checking identity, action, camera direction, and timing preservation; the final Ref2VA final cut likewise removes --turbo。

Before each run, first record one "aesthetic hypothesis" and one "biggest risk." Passing technical specifications does not equal passing creatively; keep failed runs as research samples, and in the next round change only the most likely root cause.

Delivery format for each plan

Output in the following order, and explicitly write "none" for missing items:

1. One-sentence creative and success criteria.
2. Workflow, frame size, duration, resolution, steps, and whether it currently belongs to Turbo sampling or Base final output.
3. Reference material mapping table and command-line order.
4. The complete Prompt that can be saved directly as a file.
5. Executable opc video command or dry-run command.
6. Focus of the first round of testing, failure criteria, and which single variable to change in the next round.

Completion criteria

• The media sequence numbers in the Prompt and command are completely consistent.
• Reference responsibilities do not conflict with each other.
• Sound, text, camera, and action all have clear timing or hold rules.
• The generation result passes actual frame extraction, audio track, and decoding checks.
• If two targeted iterations converge on the same defect, report the boundary and provide a plan for shot splitting, mask compositing, or post-processing, rather than continuing to blindly burn compute.