# MiniMax H3 — Prompt Guide

**Role:** turn the user's idea into a MiniMax H3 prompt. Ask at most one clarifying question, then
output the prompt in a fenced block. Verified Aug 2026; host limits override this file.

H3 is omni-modal: text/image/video/audio in one context, out comes 2K video **with native stereo
audio in the same pass**. No negative-prompt field. It fills gaps confidently — structure buys
obedience, not prettier pictures.

## Constraints

| Item | Value |
|---|---|
| Output | 2K, 24fps, AAC stereo |
| Duration | 4–15s integers (many hosts: 5–10s text/image, 5–15s reference, default 8) |
| Ratios | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, `adaptive`. **Text-to-video rejects `adaptive`** |
| Prompt | ≤7,000 chars |
| References | ≤9 images, ≤3 videos (2–15s ea), ≤3 audio, ≤12 files. **Audio can't ride alone** |

## Pick the mode — set by what's attached

| Attached | Mode | Prompt carries |
|---|---|---|
| nothing | text-to-video | everything |
| the literal opening frame (± closing frame) | first/last frame | the *motion between* frames, not their appearance |
| anything used as a reference | reference-to-video | one job per asset + the invented scene |

Pick **one** contract: text-to-video · first/last frame · omni reference (identity/wardrobe/props/
location) · performance transfer (motion only) · voice transfer · targeted edit (one clip is master).
Mixing two is where reference work falls apart.

## Six blocks

| # | Block | Omit it → |
|---|---|---|
| 1 | **Style contract** — medium, texture, palette, era, what must not drift | generic render, drifts by 6s |
| 2 | **Timeline** — time slices, one change each, each with an end state | one idea stretched thin |
| 3 | **Camera** — the one move, or refusals | unrequested drift and reframing |
| 4 | **Audio** — every sound, in order, with entry times | random room tone |
| 5 | **Text** — literal strings in quotes + treatment + position | letter-shaped gibberish |
| 6 | **Negatives** — what to refuse | soft dissolves, invented captions |

Blocks 5–6 are free. Prompt length should track how much you *didn't* hand to a reference.

## Rules

**Timeline** — `[0-3 seconds]…[3-7 seconds]…`, consecutive, non-overlapping. One primary change per
beat (two collapse into the easier one). End each beat with a state a viewer could point at ("the
bench is empty", "the wrench still in his right hand") — the end state does the work. ~4s for a prop
change or hand-off, 3s for a camera change. **The last beat gets compressed** — put the shot you care
about in the middle. Single continuous action → plain prose, no brackets.

**Camera** — it defaults to movement. To hold a frame, list the refusals: *"Locked-off static wide.
No push in, no handheld, no zoom, no dolly. The frame never moves."* When you want a move, name one
and attach a visible change ("push in to a close-up of the espresso stream"), not a bare film term.

**Audio** — own `Audio:` line. Sounds in the order they happen, with times. Music as a cue sheet
(instrumentation + structure: "kick at 3s, bass groove at 6s, one tense chord holds the last two").
Close with bans: `No music.` Quiet scenes come back genuinely quiet — ride gain in post.

**Text** — **if a word must be readable, type it.** Described text still renders, just not *your*
words. Name weight/case/family and position. Always append: *"Do not misspell it, do not add other
text, do not add subtitles."* For credits, add: each name and role appears exactly once.

**Performance** — write what a camera sees, not emotion. Not "she looks anxious" → "gaze fixed on
the floor, fingers gripping the violin neck, shoulders held up, one held breath."

**Negatives** — plain sentences in the body. Best against what the model adds unprompted: on-screen
text, camera movement, dissolves, extra people, watermarks. Weak at removing an implied subject —
describe an empty place instead of banning the person. Name real failure modes (identity drift,
wardrobe swap, broken eyelines), not "low quality". Also your main tonal control: banning fangs and
jump scares is what keeps a whimsical brief out of horror.

**References** — tokens `Image1`, `Video1`, `Audio1` (no space); the number is upload position, so
fix order first and never renumber. One job per asset, stated in text. Rule out unwanted parts by
name ("ignore the grey studio backdrop"). Name preserved features **in words as well as showing
them** — faces and hair hold, **wardrobe drifts**, so describe the garment too. References carry
identity and style; text carries the scene. Don't change pose, outfit, and angle in one generation.

**First/last frame** — prompt the motion between, not the frames. Similar aspect ratios or the
transition gets ugly; the image sets output shape. Mutually exclusive with reference images. Pinning
the last frame kills the most common cause of re-runs (the ending drifted).

**Editing** — explicit substitution list, each change paired with what stays stable. That's what
makes it localized instead of a full regeneration.

**What's load-bearing** (A/B tested): timed beats with end states, camera refusals, spelled-out text,
observable behaviour, first+last frame. **Free but marginal**: reference role labels, negative lists
— write them anyway, a sentence is cheaper than a re-run.

## Templates

**Simple shot**
```
<Subject> <action> in <environment>, at <time of day>.
<Light direction, materials, colour, texture>.
Begin on <framing>, then <the one camera move>.
Audio: <sounds in the order they happen>. No music.
```

**Timed beats**
```
<Setting and overall event.> <Style contract: medium, palette, texture, era.>

[0-3 seconds] <Opening shot. One change.> End with <pointable state>.
[3-7 seconds] <Next shot. One change.> End with <pointable state>.
[7-10 seconds] <Final shot. One change.> End with <closing state>.

Camera: <one move, or the refusals>.
Audio: <bed>, then <event sounds with times>.
Text: a card reads "<EXACT STRING>" in <weight, case, family>, <position>.
Do not misspell it, do not add other text, do not add subtitles.
Avoid: <real failure modes>.
```

**Full reference brief** — lock stable truths above the shot list; finishing sections below it.
```
[REFERENCE USE]
Image1 = <job>. Keep <features>. Ignore <what must not travel>.
Video1 = motion only. Copy <motion>. Do not use its actors, clothing, or location.
Audio1 = <name>'s voice. Keep lip sync accurate to the line below.

[LOCKS] Exactly <N> people, no extras. <Name> = face from Image1, wears <garment in words>.
        Hero prop is <object>, stays in <whose hands> until <handoff>.
[SCENE] <Location, time of day, light direction.>
[DIALOGUE] <Name>: "<exact line>"
[GEOGRAPHY] <Who is left, who is right, where each pair of eyes goes.>

[SHOT LIST]
[0-4 seconds] <framing, camera, one action>. End with <state>.
[4-8 seconds] <framing, camera, one action>. End with <state>.

[ACTING] <hands, gaze, posture, breath>
[LIGHT] <direction, contrast, palette, lens, grain>
[CAMERA] <one move, or refusals>
[SOUND] <bed, events in order, bans>
[NEGATIVES] <identity drift, wardrobe swaps, extra people, broken eyelines, subtitles>
```

## Before generating

Mode + one contract chosen · every asset has one job and its unwanted parts named · every identity,
voice and prop has exactly one owner · beats consecutive and long enough · last beat expendable ·
camera specified · performance as behaviour · every readable string typed out · audio in order ·
negatives are real failure modes · explicit ratio if text-to-video · under 7,000 chars.

**Economics:** every generation is 2K, so a bad clip costs the same as a good one and cost scales
with length. Prove the look and sound at 5s, then re-run at full length.

## Notes

Turn off auto prompt-enhancers once you've written a structured brief. API is async: POST → poll
`task_id` (~10s) → download `content.url`; `content[]` items typed `text`/`image_url`/`video_url`/
`audio_url` with roles `first_frame`, `last_frame`, `reference_image`, `reference_video`,
`reference_audio`, `base_video`. The `h3_context_ir` endpoint returns an enriched prompt without
generating. Availability, open-weight licensing, and content policy (real people, known IP) vary by
territory — check before commercial use.