Install the MiniMax H3 prompt-writing skill for agents

The MiniMax H3 prompt-writing skill gives an AI agent a defined way to turn a creative request into an H3-ready audiovisual prompt. Instead of asking the agent to improvise a format, you install one skill, identify the generation mode, and receive the fields and reference labels expected by that mode.

This tutorial is about setting up and operating the skill, not collecting every possible H3 prompting technique. It follows the files in MiniMax's official MiniMax-H3 repository, including the h3-prompt-writing skill, its base-mode guide, and its full-reference guide. For broader advice about reference roles, preservation rules, and production prompts, keep the existing MiniMax H3 prompt guide open separately.

Install the MiniMax H3 prompt-writing skill

Run the single-skill install command published in MiniMax's repository from the project where you use your AI coding or creative agent:

npx skills add https://github.com/MiniMax-AI/MiniMax-H3 --skill h3-prompt-writing

The first command installs only h3-prompt-writing. The repository currently bundles that prompt-writing skill alongside eight style-specific video-generation skills, so selecting the skill by name keeps this setup focused.

The installed skill has two reference guides:

• references/base-en.txt covers T2VA, I2VA, FL2VA, and L2VA.
• references/ref-en.txt covers full-reference Ref2VA.

That split matters more than the topic of the clip. The agent should choose a guide from the request's input relationship, then preserve that guide's field names, section order, labels, and timing notation.

Choose the H3 task mode before asking for a rewrite

Give the agent the mode explicitly whenever you know it. If you do not, describe which boundary frames and reference assets you have so it can classify the request.

T2VA, I2VA, FL2VA, and L2VA are the skill's base modes. They share one three-field body. I2VA, FL2VA, and L2VA add an exact alignment instruction above that body; T2VA begins directly with the first field.

Ref2VA is not just the same template with more attachments. It has a separate six-section contract for defining references, classifying their role, analyzing retention, and describing when each reference takes effect.

Do not choose Ref2VA merely because a request is detailed. Choose it when referenced content has to be defined and tracked across the output.

Know the exact base-mode output structure

For every base mode, the body uses these three field names in this order:

integrated_multimodal_description: [Shot 1] ...

overall_soundscape: ...

non_diegetic_music: ...

integrated_multimodal_description carries the visual style, composition, subjects, environment, actions, camera, dialogue, singing, and synchronized diegetic sound along the timeline. overall_soundscape summarizes ambience, action sounds, and non-verbal human sounds across the full clip. non_diegetic_music describes music heard by the audience but not by characters; it is N/A when there is none.

The alignment line changes with the keyframe mode. The official guide specifies these forms:

I2VA
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.

FL2VA
How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot N) aligns with the S.SS-second mark of the target video.

L2VA
How the reference pictures align with the target video — <Picture 1> (from [Shot N]) aligns with the S.SS-second mark of the target video.

In the finished rewrite, N becomes the actual final shot number and S.SS becomes the effective duration with exactly two decimal places. The line comes first, followed by one blank line and the three-field body. T2VA has no alignment line.

Shot timing has its own convention: [Shot 1] has no timestamp, while later shots begin with a strictly increasing cut time such as [Shot 2] At 00:03.500, .... A base-mode output is ready only when the chosen mode, alignment line, final shot, and duration agree.

Know the exact Ref2VA output structure

Full-reference mode must return six sections in this order:

subject_definitions:
...

summary:
[reference generation] ...

retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - ...

detailed_description:
...

overall_soundscape:
...

non_diegetic_music:
...

Each label has a distinct job:

• <Subject N> names reusable visible content, such as a person, object, scene, style, action, interface, or effect.
• <Picture N> names an image used as a concrete target frame, keyframe, composition anchor, or storyboard reference.
• <Video N> names a source video used for editing, continuation, or whole-video temporal structure.
• <Audio N> names a standalone audio signal or an enabled synchronized audio track that is copied or referenced.

These categories are numbered independently. A visual subject extracted from a reference video still uses <Subject N>; <Video N> represents the asset or its whole-video relationship. Likewise, a video file containing sound does not automatically create an <Audio N> entry.

The summary begins with one or more official task types inside square brackets: keyframe completion, reference generation, video editing, video continuation, audio reuse, or audio reference. Multiple applicable types join with +.

retention_analysis then gives every label an explicit relationship. Visible content uses fully_preserved, partially_preserved, attribute_transfer, or weak_reference. Audio uses fully_copy, partially_copy, reference, or weak_reference. The marker should describe the intended relationship, not serve as a general quality score.

Finally, Ref2VA replaces the base field integrated_multimodal_description with detailed_description. For generation tasks, MiniMax's guide normally calls for 350–500 English words here, with enough detail to establish each shot's composition, subjects, environment, lighting, actions, state changes, camera movement, sound, and reference usage. Dialogue-dense work prioritizes fitting the complete spoken timeline instead of mechanically reaching that range.

Ask the agent for a base-mode prompt

After installation, call the skill by name and provide a compact production brief. The skill's bundled agent configuration uses $h3-prompt-writing in its default request.

Here is a pasteable FL2VA request:

Use $h3-prompt-writing to rewrite this request as a MiniMax H3 FL2VA prompt.

Duration: 8.00 seconds.
Picture 1: the supplied first frame, a closed field notebook on a rain-darkened café table.
Picture 2: the supplied last frame, the same notebook open to a pressed violet flower.
Action: one hand enters, opens the notebook, and stops on the flower page.
Camera: one continuous shot with a slow, small push in.
Sound: quiet café room tone, rain against glass, paper and cover movement.
Dialogue: none.
Non-diegetic music: none.

Preserve the table position, notebook identity, hand continuity, lighting direction, and final composition. Return only the final rewrite in the official base-mode structure.

The useful parts are the explicit mode, exact duration, clear role for each picture, achievable transition path, sound decisions, and output constraint. The agent should return the FL2VA alignment line first, use the three shared fields, and land Picture 2 at 8.00 seconds.

For I2VA, replace the second image with the actions that develop from the first frame. For L2VA, describe the intended event before the supplied final frame. For T2VA, remove picture references and give the agent enough information to construct the opening composition and audiovisual sequence from text.

Ask the agent for a full-reference prompt

Ref2VA requests work better when every asset has a stated role. Here is a pasteable agent request that separates identity, motion, and audio guidance:

Use $h3-prompt-writing to rewrite this request as a MiniMax H3 full-reference Ref2VA prompt.

Target: a 10-second, three-shot product demonstration.
Picture 1: preserve the reusable bottle identity, silhouette, cap, label placement, and material finish.
Video 1: reference only the demonstrator's hand action and the three-shot cut rhythm; do not reuse its product, set, or wardrobe.
Audio 1: reference the percussion tempo for edit timing without copying the source signal.

Shot flow:
1. Establish the bottle centered on a stone counter.
2. A hand presses the pump once and the lotion lands on the back of the other hand.
3. End on the bottle and a small lotion smear beside it.

No dialogue or visible text beyond the preserved label. Use new room ambience and restrained audience-only percussion. Return only the six official Ref2VA sections in their required order, with consistent labels and explicit retention markers.

The request does not attempt to write subject_definitions for the agent. It gives enough provenance and reuse intent for the skill to construct those definitions, classify the summary as reference generation plus audio reference, and explain the expected preservation or transfer for each label. If an asset should affect only one property, say so: a motion clip can guide hand action without also controlling the performer, wardrobe, set, product, camera, or audio.

Use a two-pass agent workflow

Treat the skill output as a structured draft with a deterministic review pass.

1. Write the source manifest. List every text instruction, frame, image, video, and audio asset. State what each asset should and should not control.
2. Lock the task mode. Choose T2VA, I2VA, FL2VA, L2VA, or Ref2VA before the rewrite.
3. Give the agent delivery constraints. Include duration, shot count, boundary-frame timing, dialogue, visible text, and whether music exists.
4. Request only the official structure. This makes missing fields, reordered sections, and commentary outside the prompt easy to detect.
5. Review without changing the brief. Check structure, references, timing, spoken text, and audio placement first.
6. Revise one layer at a time. Ask the agent to correct a specific mismatch while preserving the accepted sections.

For the second pass, use a targeted instruction:

Review the rewrite against the h3-prompt-writing guide. Keep the creative brief unchanged. Correct only field order, unresolved reference labels, shot timestamps, duration alignment, dialogue preservation, and diegetic versus non-diegetic audio placement. Return the corrected prompt only.

This review prompt limits the chance that a format correction silently becomes a creative rewrite.

Review the output before generation

Use this checklist after every agent rewrite:

• Mode: The output uses the guide for the actual input relationship.
• Base structure: T2VA starts with integrated_multimodal_description; I2VA, FL2VA, and L2VA place the correct alignment instruction above it.
• Ref2VA structure: All six fields appear once and in the required order; detailed_description is used instead of integrated_multimodal_description.
• Labels: Every <Subject N>, <Picture N>, <Video N>, and <Audio N> keeps one meaning across all sections. No undefined label appears later.
• Reference roles: Subjects, concrete frames, whole-video relationships, and audio signals use the correct label type.
• Retention: Every tracked reference has an allowed relationship marker consistent with the request.
• Shots: [Shot 1] has no timestamp. Later cut times strictly increase and remain within the requested duration.
• Boundary frames: I2VA begins from the first frame, FL2VA reaches the last frame, and L2VA converges on its supplied final frame.
• Speech and text: Dialogue, lyrics, and visible scene text keep their original language and wording. Rewrite prose remains in English.
• Audio: Dialogue and synchronized diegetic events stay in the main description; full-clip ambience belongs in overall_soundscape; audience-only score belongs in non_diegetic_music.
• Feasibility: The described actions fit the duration, and the final moment is resolved rather than cut off unintentionally.
• Output hygiene: The agent has not added a plot summary, an unsupported asset, a conflicting role, or commentary outside the requested prompt.

Move from the agent draft to OmniArt

The skill is most useful as a preparation layer: it converts an asset manifest and creative brief into a prompt you can inspect before generation. Keep the accepted output with its source files, because its labels only stay useful when the asset order and meaning remain stable.

Then bring the reviewed prompt and matching assets into OmniArt's video creation workspace. If you need to improve the creative brief itself rather than its H3 field structure, return to the MiniMax H3 prompt guide. The clean division is simple: use the guide to decide what you want, use h3-prompt-writing to express it in the selected H3 contract, and use the checklist to stop structural mistakes before they reach a generation run.