Hailuo (海螺) prompt builder that emits MiniMax's 15 documented camera commands — [推进], [变焦推近] and the rest — plus the H3 grammar.
Hailuo Prompt Builder
MiniMax publishes a formal camera command language for the Hailuo 2.x line — fifteen commands, written in square brackets — and its own documentation says that free-form descriptions also work, but explicit commands yield more accurate results. This tool is a compiler for that grammar: pick the moves, and it assembles a bracketed prompt on the beats you chose. MiniMax-H3 (海螺 3.0), released 31 July 2026, uses an entirely different grammar, so the model selector switches the whole form. Everything is built in your browser; no model is called.
Do this first
prompt_optimizer defaults to true — Hailuo rewrites your prompt first
MiniMax's own wording: "Whether to automatically optimize the prompt. Defaults to true. Set to false for more precise control." So under the default the model does not see what you wrote — it sees Hailuo's rewrite of it, and the structure you arranged, along with the command you placed on a particular beat, can be flattened on the way. Switch it off for a structured prompt to mean anything.
"prompt_optimizer": false
Shot beats and camera moves
Write what happens beat by beat, then pick the camera move for that beat. Commands in one bracket fire simultaneously (three at most is the recommendation); separate brackets fire in order. The bracket lands at the end of its beat — the moment the move happens. Empty beats are skipped.
- [Push in]
- The camera body actually travels forward. Foreground and background shift against each other, space compresses, and the viewer feels themselves approach.
- [Zoom in]
- Only the focal length lengthens. The image magnifies with no parallax at all, which reads flatter and more like surveillance footage.
Widely circulated Chinese term tables collapse both into one "推镜头 → Zoom in" row. They are different shots.
Request parameters — do not put these in the prompt
Duration and resolution are the API's duration and resolution fields. Putting them in the prompt only spends characters. Set them in the Hailuo UI or your request.
Shots
Shot 1 carries no timestamp; every later shot opens with an absolute cut time in MM:SS.mmm and must strictly increase. MiniMax's guidance: a cut should introduce new information — if only the distance or angle changes, prefer camera motion over a cut.
Camera motion (a sentence, not a label)
H3's guide asks for camera motion inside the action sentence, as type plus amplitude plus speed — and says medium amplitude and normal speed are usually omitted. The composed phrase is inserted into whichever shot you last edited.
Dialogue
Speakers are tagged (S1), (S2) and stay stable across shots. Only the language tag and the verbatim line go inside <d>…</d>; the speaker, the action and the delivery stay outside it.
Left empty, these are written as N/A — the sentinel the guide specifies.
Request parameters — do not put these in the prompt
Duration and resolution are the API's duration and resolution fields. Putting them in the prompt only spends characters. Set them in the Hailuo UI or your request.
Tip: this builder only compiles text. Copy the result into Hailuo (海螺 / MiniMax) yourself — no model is called and nothing is sent anywhere.
How this Hailuo prompt builder works
Turn the prompt optimiser off first
MiniMax's prompt_optimizer defaults to true — Hailuo rewrites your prompt before the model ever sees it. The structure you were careful about, and the camera command you placed on a particular beat, can both be rewritten away. For a structured prompt to pay off at all, send prompt_optimizer: false in the API call, or switch the optimiser off in the web UI. Almost no English-language guide mentions this, and everything below depends on it.
Pick the grammar: 2.x or H3
Hailuo 2.3 / 02 / T2V-01-Director read bracketed camera commands. MiniMax-H3 (海螺 3.0), released 31 July 2026, uses an entirely different grammar — named fields, [Shot N] blocks with absolute cut times, and camera motion written as a sentence rather than a label. Nothing in MiniMax's documentation supports putting 2.x brackets into an H3 prompt. Choose the model above and the whole form switches.
Put the camera command on the beat where the move happens
In 2.x mode, write what happens beat by beat and pick the camera move for that beat. Commands inside one bracket fire simultaneously (MiniMax recommends no more than three); separate brackets fire in prompt order. The bracket belongs at the moment of the move, not appended at the end — which is exactly how MiniMax writes its own example: "A man picks up a book [Pedestal up], then reads [Static shot]."
Read the counter and the warnings, then copy
The prompt ceiling is 2,000 characters (H3's content text caps at 7,000) and the counter sits above the output. Problems are listed live underneath: more than three commands in one bracket, [Static shot] combined with a move, a 2.x bracket smuggled into an H3 prompt, a cut time that does not increase. Duration and resolution are request parameters rather than prompt text, so they sit in their own panel for you to set in the Hailuo UI or your API call.
How this Hailuo prompt builder works
Hailuo has a formal camera language, and most builders emit none of it
Most text-to-video prompting advice teaches the same thing: write a more detailed description. That is not wrong, but it misses what actually makes Hailuo different. MiniMax publishes a formal camera command language for the 2.x line — fifteen commands, written in square brackets, mixed straight into the prompt. Its documentation puts the case plainly inside the prompt field's own description: free-form descriptions also work, but explicit commands yield more accurate results. In other words, a prompt tool that offers only a free textarea is teaching the path the vendor itself calls worse. This builder turns those fifteen into buttons and assembles them by the documented rules: commands inside one bracket fire simultaneously, with three the recommended maximum; separate brackets fire in prompt order; and a bracket belongs on the beat where the move happens, not appended at the end.
The most misunderstood pair in that vocabulary is [Push in] and [Zoom in]. In Chinese both are commonly called 推镜头, and widely circulated term tables merge them into a single row mapped to "Zoom in". They produce completely different footage. [Push in] moves the camera body forward: foreground and background shift against one another, space compresses, and the viewer gets a bodily sense of approaching. [Zoom in] only lengthens the focal length — the image magnifies with no parallax whatsoever, reading flatter and more like surveillance material. Hitchcock's vertigo effect exists precisely because the two are different, run against each other at once. This tool shows them side by side with a line of gloss each: a correctness fix disguised as a UI.
"Free-form descriptions also work, but explicit commands yield more accurate results." — MiniMax, in its own prompt field documentation.
Two Hailuos, two grammars, and the switch that is on by default
One switch decides whether any of the above survives: prompt_optimizer. MiniMax documents it as defaulting to true, describes it as automatically optimising the prompt, and notes that setting it to false gives more precise control. Under the default, then, the model does not read what you wrote — it reads Hailuo's rewrite of it, and the sections you arranged and the [Static shot] you placed on the second beat can be smoothed away in that step. For a casual clip it is a thoughtful feature. For a structured prompt builder it is existential, which is why it sits at the top of this page — and why almost no English-language guide mentions it at all.
Finally, "Hailuo" now names two things. Hailuo 2.3 and 02 read the brackets. MiniMax-H3 (海螺 3.0), released 31 July 2026, replaced the grammar wholesale: three named fields (integrated_multimodal_description, overall_soundscape, non_diegetic_music), [Shot N] blocks carrying absolute MM:SS.mmm cut times, speakers tagged (S1), dialogue wrapped in <d>[Language] …</d>, and camera motion written as a sentence inside the action — type, then amplitude, then speed, with medium amplitude and normal speed usually left out. Its guide adds two rules worth stealing for any model: prefer concrete visual and audio detail over abstract words like "cinematic", and let a cut introduce new information — if only the distance or angle changes, use camera motion instead. A builder pretending there is one Hailuo is already wrong on step one, so the selector here switches the entire grammar rather than a single dropdown value.
There is no orbit command — any tool offering one is inventing it
MiniMax publishes a formal camera command language for the Hailuo 2.x line — fifteen of them: Truck left, Truck right, Pan left, Pan right, Push in, Pull out, Pedestal up, Pedestal down, Tilt up, Tilt down, Zoom in, Zoom out, Shake, Tracking shot, Static shot.
The documentation says plainly that free-form descriptions also work, but explicit commands yield more accurate results. Most prompt tools offer only a free textarea, which teaches the path the vendor itself calls worse.
Commands inside one bracket fire simultaneously ([Pan left,Pedestal up]); MiniMax recommends no more than three. Commands in separate brackets fire sequentially, in the order they appear in the prompt.
[Push in] moves the camera body — parallax changes and depth compresses. [Zoom in] changes focal length only, with no parallax at all. Widely circulated Chinese term tables merge both into one "推镜头 → Zoom in" row. They are different shots.
There is no [Orbit] command. It is not among the fifteen, and any tool offering an orbit button invented it.
prompt_optimizer defaults to true: unless you switch it off, Hailuo rewrites your prompt before the model sees it. A structured prompt only pays off with the optimiser disabled.
The prompt ceiling is 2,000 characters — stated on all four /v1 mode pages. MiniMax-H3's /v2 endpoint raises the content text cap to 7,000.
Duration and resolution are request parameters, not prompt text. Hailuo 2.3 and 02 accept 6 or 10 seconds only, and 1080P is available at 6 seconds; H3 takes any integer from 4 to 15.
MiniMax-H3 (海螺 3.0) launched on 31 July 2026 with a different grammar entirely: the named fields integrated_multimodal_description, overall_soundscape and non_diegetic_music, [Shot N] blocks carrying MM:SS.mmm cut times, and camera motion written as type plus amplitude plus speed.
This tool assembles the prompt entirely in your browser — your input is never uploaded, never sent to a model, and never stored.
Frequently Asked Questions
- No. It compiles what you type into a prompt in the grammar MiniMax publishes, entirely in your browser. It does not call Hailuo, MiniMax or any model, and does not go online. To generate a video you copy the prompt and run it in Hailuo yourself.
- If you care about structure, yes. MiniMax's documentation states it defaults to true, that it automatically optimises the prompt, and that setting it to false gives more precise control. Leave it on and the sections you arranged, and the camera command you placed on a particular beat, can be rewritten before the model sees them. For a casual clip it is fine; for a structured prompt to mean anything, switch it off.
- Fifteen: Truck left, Truck right, Pan left, Pan right, Push in, Pull out, Pedestal up, Pedestal down, Tilt up, Tilt down, Zoom in, Zoom out, Shake, Tracking shot, Static shot. The Chinese and English documentation lists map one to one. There is no Orbit.
- [Push in] physically moves the camera forward: the relationship between foreground and background changes, space compresses, and the viewer feels themselves approach. [Zoom in] only lengthens the focal length — the image magnifies with no parallax at all, which reads flatter and more surveillance-like. Many Chinese term tables translate both as 推镜头 / Zoom in, which is why this tool shows them side by side.
- Commands inside one bracket, comma-separated, fire simultaneously — [Pan left,Pedestal up] — and MiniMax recommends at most three per bracket. Brackets in different places in the prompt fire in order. So "push in, then hold" is two brackets on two beats, not [Push in,Static shot], which is a self-contradictory simultaneous pair the tool will warn you about.
- Because Hailuo has no such parameter. MiniMax's documented request parameters for the video API are model, prompt, prompt_optimizer, fast_pretreatment, duration, resolution, callback_url and aigc_watermark — negative_prompt is not among them. Some third-party wrappers show a negative box, but that is their own pre-processing. Describe what you do want instead; that is what the model actually reads.
- No. H3 uses named fields (integrated_multimodal_description, overall_soundscape, non_diegetic_music), [Shot N] blocks with MM:SS.mmm cut times, and camera motion written as type plus amplitude plus speed inside the action sentence — its guide explicitly says not to stack motion as labels. Nothing in MiniMax's documentation says H3 reads 2.x brackets, and [Shot N] is itself meaningful H3 syntax, so this tool warns when a 2.x bracket appears in an H3 prompt.
- No. They are the duration and resolution request parameters; putting them in the prompt just spends characters. Hailuo 2.3 and 02 take 6 or 10 seconds, with 1080P at 6 seconds only; T2V-01-Director is 6 seconds at 720P; H3 takes any integer from 4 to 15 at 768P or 2K. The tool keeps them in their own panel so you can copy the settings across.
- No. MiniMax's own formula is short: subject plus scene plus motion for a basic prompt, then camera movement and mood for a precise one. The Chinese prompting community says the same thing — you do not need all eight components. Empty fields are skipped entirely, so the prompt never carries a heading with nothing under it.
- Because the two endpoints have different ceilings. MiniMax documents the prompt field on the /v1 video endpoint — the one Hailuo 2.3, 02 and the Director models use — as a maximum of 2,000 characters; the /v2 endpoint that MiniMax-H3 uses caps the content text at 7,000. The counter reads the mode you have selected and warns when you cross the line, because anything past it is cut off rather than read.
Related News
You may be interested in these recent stories from our newsroom.
-
The Evidence That AI Wrote PhantomRaven Is Its Comments
CrowdStrike calls the stealer almost certainly machine-generated, on redundant commenting, a leftover placeholder URL and token statistics.
-
Claude Opus 5 Wrote the Exploit. Three People Aimed It.
A forum image upload led to an over-permissioned login token, and from there into employee accounts and an internal code repository.
-
Google Gemini Breached Three Real Companies During a Security Test
A capture-the-flag exercise handed a model internet access it was never meant to have, aimed at a company that turned out to exist.
Method & sources
How it computes
The form is compiled into a prompt in the grammar of the model you pick, and the two families are not interchangeable: the 2.x line (MiniMax-Hailuo-2.3 / 02 / T2V-01-Director) takes bracketed camera commands inside otherwise free-form text, while MiniMax-H3 takes a structured block with named fields, [Shot N] cut times in MM:SS.mmm and camera motion written as prose inside the action sentence. Pure string assembly — nothing is sent anywhere and no model is called.
What this tool implements
- Fifteen camera commands, exactly as MiniMax documents them in the prompt field description, with the Chinese and English forms mapped one to one. There is NO [环绕] / [Orbit] command — it is not in either language's list, and a tool that offers one invented it.
- 推进 and 变焦推近 are DIFFERENT SHOTS and are kept apart here. Widely circulated Chinese term tables collapse both into one "推镜头 → Zoom in" row. Moving the camera changes parallax and compresses depth; changing focal length does neither.
- Simultaneous commands share one bracket, comma-separated; MiniMax's own rule is 建议组合不超过 3 个 (no more than three recommended), which the tool shows as a warning rather than enforcing. Separate brackets fire in prompt order.
- As of 18 August 2026 the vendor's English page's sequential example used "[Push out]", which is not in its own command list; the Chinese original says [拉远]. This tool treats [Pull out] as canonical.
- The two control surfaces are never mixed. [Shot N] is a meaningful token in H3's grammar, so emitting a 2.x bracket into an H3 prompt is not a style choice — at best it is noise the model must ignore.
- Character ceilings are enforced per endpoint: 2,000 for the /v1 2.x modes, 7,000 for H3 on /v2. prompt_optimizer defaults to true on the 2.x endpoints and the page tells the reader to switch it off.
Sources
- MiniMax platform documentation, video generation API (text-to-video): https://platform.minimaxi.com/docs/api-reference/video-generation… — with the i2v, fl2v, s2v and v2-create pages. The 15-command list, the ≤3-per-bracket recommendation, the 2,000-character prompt maximum and prompt_optimizer (default true) all sit inside the prompt field description; there is no separate camera guide. Re-read 13 September 2026.
- MiniMax-H3 prompt-writing guide: https://github.com/MiniMax-AI/MiniMax-H3 under skills/h3-prompt-writing/ — the guide shipped with the open weights, so it is the vendor's own guidance rather than a community reading. MiniMax-H3 (海螺 3.0) launched 31 July 2026 on /v2/video_generation. Directory confirmed present 13 September 2026.
What can make this go out of date
- MiniMax revises these docs and ships new models — H3 arrived in July 2026 with an incompatible grammar. The command list, the character ceilings and the endpoint paths are snapshots (18 August 2026, re-checked 13 September 2026) and nothing refetches them.
- If the vendor corrects the "[Push out]" error in its English documentation, the note above becomes history rather than a correction, and should be re-checked before it is repeated.
Pick up where you left off
Stored only in this browser — never sent to our servers.