# MiniMax H3 Prompt Studio

> Nine production workflows from the MiniMax-H3 skills repository, run over one creative brief and
> ending in MiniMax H3 structured video prompts. Includes an offline H3 prompt-format validator that
> needs no account, and a per-run model picker so you choose which model writes the package.

URL: https://minimax-h3-prompt.skillsafe.ai/
API: https://minimax-h3-prompt.skillsafe.ai/api.html

## What it does

Takes one creative brief and produces the pre-production package a MiniMax H3 production actually
needs: project briefs, product fact sheets, provenance manifests, story outlines, character and
scene card prompts, six-column shot tables, beat storyboards, image prompts, and H3 video prompts
in the published prompt structure.

## What it does not do

It does not generate video, and it says so. Eight of the nine upstream skills declare
`compatibility: Requires the MiniMax Hub agent` and call Hub tools for image, video, speech and
editing. Their own rule for a runtime without those tools is to downgrade the deliverable to a
pre-production package rather than claim a finished video. That is what this app produces. It can
render a preview image for the workflows whose approval gate is an image, using the platform's
image model, and that is the only asset it generates.

## The nine workflows

| Lane | Source skill | Produces |
| --- | --- | --- |
| h3-prompt | h3-prompt-writing | An H3 prompt in T2VA, I2VA, FL2VA, L2VA or Ref2VA structure |
| handdrawn | handdrawn-live-video-generator | A 15-second 16:9 prompt in eleven fixed paragraphs |
| co-op-intro | co-op-game-intro-generator | A sixteen-field confirmation-image prompt, then the video prompt |
| product-ad | minimalist-product-ad-generator | Product facts, spine, copy, three anchors, an eleven-column beat table |
| brand-promo | brand-promo-video-generator | Brand truth sheet, provenance manifest, frame-aware beat plan |
| mv-subtitle | music-video-subtitle-generator | Pre-flight lock, locked lyrics, modular multi-shot MV script |
| paper-collage | paper-collage-explainer-generator | Production plan, still specs and stop-motion assembly prompts |
| papercraft | papercraft-stop-motion-explainer | Style DNA, diorama staging, prompt set and storyboard |
| anim-short | 3d-animation-short-generator | Brief, outline, character and scene cards, six-column shot table |

Seven of the nine run in two stages, `plan` then `build`, with the boundary placed at the approval
gate the upstream skill itself defines. The `task` field in the API is `<lane>:<stage>`, plus the
bare `route` task which recommends a workflow from a plain-language request.

## The H3 prompt structure it writes

Base modes (T2VA, I2VA, FL2VA, L2VA) produce a mode-specific alignment instruction, a blank line,
then `integrated_multimodal_description`, `overall_soundscape` and `non_diegetic_music` in that
order. Full-reference mode (Ref2VA) produces `subject_definitions`, `summary`, `retention_analysis`,
`detailed_description`, `overall_soundscape` and `non_diegetic_music`.

Shots are marked `[Shot N]`; the first carries no timestamp and later ones open `At MM:SS.mmm,`.
Camera motion is drawn from a closed vocabulary with `with small amplitude` / `with large amplitude`
and `at slow speed` / `at fast speed`. Speakers get stable `(S1)` ids and dialogue is wrapped
`<d>[Language] ...</d>` with the original words preserved. Reference labels are `<Picture N>`,
`<Video N>`, `<Audio N>` and, in full-reference mode, `<Subject N>`.

## The free validator

The page ships a client-side validator for that structure. It runs with no account and no token,
against the prompt the model GENERATES - not against what you type, which is a description of the
video rather than a prompt. It checks field presence and order, the mode-specific instruction line and its two-decimal
duration, shot numbering and cut-time monotonicity, the camera vocabulary, speaker id ordering,
dialogue wrapper wellformedness, the voiceover lips-closed rule, reference-label resolution, the
Ref2VA task-type prefix and relationship markers, and the sentence budgets on the two audio fields.
Its findings are listed above the generated prompt, so a rewrite that drops a reference label or
writes a cut time out of order is visible rather than plausible-looking.

## Model selection

Three independent selections: a text model for the build stages, a cheaper text model for routing,
and an image model for preview renders. Text models are restricted to the GPT-5.6 family - Sol,
Terra and Luna - so every package is written to one standard and the only variable is how much
reasoning you are paying for. The image model is exempt because no GPT-5.6 image model exists. The chosen model rides on every estimate and every run as
`$model`, so the price you are shown is the price of the model you picked.

The catalogue is a snapshot taken from `GET /v1/models` at release time, with its date shown beside
the picker. It is not read live in the browser, and the app does not claim to: `/v1/models` returns
no `access-control-allow-origin` header, so a fetch from any app origin is blocked by CORS. The live
call is still attempted on every load and will start working if that ever changes. Availability is
not taken on trust either - `/estimate` re-checks `$model` server-side every time a step is priced,
so a model retired since the snapshot prices out with the platform's own message rather than failing
after you press run.

## Sources and attribution

Every workflow is derived from a skill in https://github.com/MiniMax-AI/MiniMax-H3/tree/main/skills
and each lane links to its own upstream SKILL.md. The repository ships no licence, so this app does
not redistribute the skill text: the workflow rules are restated in our own words. What is
reproduced exactly is the functional interface - field names, table column sets, fixed paragraph
orders, controlled vocabularies - because those are specification rather than expression.

Not affiliated with or endorsed by MiniMax.

## Pricing

Free to browse, free to validate a prompt, free to get a workflow shortlist. Writing a package is a
metered run charged at actual usage on the selected model. The estimate shown before you run is a
reservation, not a price; the settled charge is usually lower and is shown afterwards.